Hochschule DarmstadtFachbereich Informatik

Efficient Biometric Identification in Large-ScalePalm Vein Databases

Abschlussarbeit zur Erlangung des akademischen GradesMaster of Science (M.Sc.)

vorgelegt von

Benedikt-Alexander Mokroß725215

Referent: Prof. Dr. Christoph BuschKorreferent: Prof. Dr. Christoph WentzelKorreferent: MSc. Pawel Drozdowski

Ausgabedatum: 01.06.2017Abgabedatum: 01.12.2017

ii

Erkärung

Ich versichere hiermit, dass ich die vorliegende Arbeit selbständig verfasst und keine anderenals die im Literaturverzeichnis angegebenen Quellen benutzt habe. Alle Stellen, die wörtlichoder sinngemäß aus veröffentlichten oder noch nicht veröffentlichten Quellen entnommen sind,sind als solche kenntlich gemacht. Die Zeichnungen oder Abbildungen in dieser Arbeit sindvon mir selbst erstellt worden oder mit einem entsprechenden Quellennachweis versehen. DieseArbeit ist in gleicher oder ähnlicher Form noch bei keiner anderen Prüfungsbehörde eingereichtworden.

Darmstadt, den 01.12.2017 Benedikt-Alexander Mokroß

iii

Abstract (English)

With the increasing number and size of biometric systems around the world, biometric databa-ses are rapidly growing. Biometric systems use measurable and distinctive human characteristicsfor the purpose of automatised recognition or identity verification tasks. This thesis aims tocontribute to the research in the area of biometric workload reduction in open-set identificationscenarios in large-scale palm vein databases. The palm vein as a biometric characteristic is verydifficult to obtain a copy from without notice of the individual and it records a steady growthin biometric market share. Therefore, it has been selected as the biometric characteristic forthis project.

The main research in this thesis was carried out using a biometric indexing approach basedon Bloom filters and binary search trees, which has already been successfully applied for theiris characteristic. To transform the extracted palm vein patterns in a Bloom filter compatiblerepresentation, an approach based on Fourier transformations has been chosen. This representa-tion was introduced for the fingerprint characteristic using defined reference points as an input.A so-called feature extraction pipeline is proposed that extracts such reference points and yieldsthem in a usable form for further processing. This system has shown an acceptable biometricperformance and a high workload reduction, as well as showing a drastic speed improvementcompared to conventional pattern affinity approaches.In addition to the Bloom filter-indexing approach, a less complex indexing approach used waspresented, merely employing the Spectral Minutiae Representation used, utilising binary searchtrees. The presented system adopts optimisation approaches presented for the Bloom filter bi-nary search trees and achieves a higher biometric performance than the Bloom filter-indexingapproach while sacrificing a small amount of workload reduction. To the best of the authorsknowledge, this is the first study of such an indexing approach using the selected representation.

It was discovered that the high fuzziness of palm vein imaging impairs the biometric per-formance of the base representation and thus also impairs the biometric performance of theindexing approaches. Facing the acceptable but not excellent biometric performance, promisingfurther research topics are presented, whereby their recorded biometric performance resultsshow promise and open new possible avenues of research in this area.

iv

Abstract (German)

Mit der international steigenden Anzahl an biometrischen Systemen, deren immer größerenausmaßen und deren zunehmenden Einsatzbereichen, resultierend aus zunehmender Akzeptanzund zunehmenden Vertrauen in die Biometrie, steigt auch die Anzahl an immer größeren, bio-metrischen Datenbanken. Biometrische Systeme nutzen mess- und bestimmbare Eigenschaftenbzw. Charakteristiken des menschlichen Körpers um dessen Identität zu bestimmen oder zu ve-rifizieren. Der Betrieb solcher enormen biometrischen Systeme benötigt außerordentlich starkeRechenkapazitäten um eine für die Benutzer akzeptable Zugriffszeit zu gewährleisten. Deswegenzielt diese Thesis darauf ab, einen Teil zu der Forschung im Gebiet der Effizienzsteigerung fürbiometrische Systeme im Identifikationsbetrieb beizutragen. Als zentrale biometrisches Charak-teristikum wurde die Handvene ausgewählt, da diese unter anderem sehr schwer unbeobachtetzu erfassen ist und der Marktanteil an Handvenensystemen stetig wächst.

Die Kernthemen der Forschung wurden mittels eines biometrischen Indizierungsverfahrensauf Basis von Bloom Filtern und binären Suchbäumen analysiert, welches bereits erfolgreichfür die Iris getestet wurde. Um die Handvenenstruktur in eine Bloom Filter kompatible Dar-stellung zu überführen, wurde eine für Fingerabdrücke entworfene Darstellung auf Basis vonFouriertransformationen ausgewählt. Für diesen Schritt wird eine eigens entworfene Bildverar-beitungspipeline vorgestellt, die die Handvenenstruktur in ein für die Fouriertransformationengeeigente Darstellung konvertiert. Die erziehlte Erkennungsleistung in den Ergebnissen ist ak-zeptabel und das System weist eine hohe Effizienz auf.Neben der Indizierung mittels Bloom Filtern wurde eine weniger komplexe Indizierungsmethodevorgestellt, welche sich nur auf die Darstellung nach den Fouriertransformationen und binärenSuchbäumen stützt. Die vorgestellte Methode nutzt Optimierungen im Suchvorgang, wie sieim Bloom Filter System eingesetzt werden, und erziehlt dabei eine bessere Erkennungsleistungbei minimal niedrigerer Effizienz als das Bloom Filter System. Diese Methode ist nach bestenWissen und Gewissen erstmalig in dieser Thesis vorgestellt und erprobt worden.

Es stellte sich heraus, dass die Fouriertransformationen mit der Unschärfe der Handvenennicht ideal funktioniert, was die Erkennungsleistung des Systems reduziert. Im Hinblick auf dieakzeptable, jedoch nicht ideale, Erkennungsleistung wurden vielversprechende weiterführendeForschungsansätze vorgestellt. Die erfolgversprechenste Methode wurde dargestellt.

v

Acknowledgements

I would like to thank my supervisors Professor Christoph Busch and Professor ChristophWentzel for making this project possible and for the excellent communication, guidance andadvice throughout the entire project period. I am especially grateful to M.Sc. Pawel Drozdowskifor the day-to-day supervision and feedback. I also thank Doctor Alexander W. Lenhardt forsupplying me with hardware, office space, time and everything I needed during the entire pro-ject. I owe my gratitude to Myriam Hilfenhaus for her moral support and her effort to providestability and harmony for me.

Portions of the research in this paper use the CASIA-MS-PalmprintV1 collected by theChinese Academy of Sciences’ Institute of Automation (CASIA) as well as The Hong KongPolytechnic University (PolyU) Multispectral Palmprint Database and the PUT Vein Databaseby the Poland Institute of Control and Information Engineering. The sources of the re-usedimages are attributed directly in the text. The data processing and visualisation in this projectwas done using the OpenCV software library.

vi

Contents

1 Introduction 11.1 Thesis Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.2 Thesis Organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Biometric System Fundamentals 32.1 Veins as a Biometric Characteristic . . . . . . . . . . . . . . . . . . . . . . . . . 32.2 Generic Biometric System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2.1 Workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.2.2 Subsystems in this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.2.3 Operation Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.2.4 Template protection requirements . . . . . . . . . . . . . . . . . . . . . . 11

2.3 Chapter Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3 Related Work 133.1 Vascular Biometric Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133.2 Workload Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3.2.1 Serial combination of algorithms . . . . . . . . . . . . . . . . . . . . . . 153.2.2 Classification, Clustering and Binning . . . . . . . . . . . . . . . . . . . 153.2.3 Indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.3 Chapter conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4 Palm Vein Feature Detection 184.1 Recommendations for high-quality vascular imaging . . . . . . . . . . . . . . . . 194.2 ROI Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204.3 Image Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4.3.1 Non-Local Means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224.3.2 Non-Linear Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

4.4 Vein Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244.5 Minutiae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254.6 Chapter Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

vii

5 Spectral Minutiae Representation 275.1 Spectral Minutiae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

5.1.1 Spectral Minutiae Location Representation . . . . . . . . . . . . . . . . . 285.1.2 Spectral Minutiae Orientation Representation . . . . . . . . . . . . . . . 295.1.3 Spectral Minutiae Complex Representation . . . . . . . . . . . . . . . . 29

5.2 Spectral Minutiae Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295.3 Normalisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315.4 Feature Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5.4.1 Column Principal Component Analysis . . . . . . . . . . . . . . . . . . . 315.4.2 Line Discrete Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . 33

5.5 Binarisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335.6 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5.6.1 Spectral Minutiae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355.6.2 Binary Spectral Minutiae . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5.7 Extending the Spectral Minutiae Representation with quality data . . . . . . . . 365.7.1 Retrieving and applying minutiae-reliability data in vascular modalities . 385.7.2 Retrieving and applying location-accuracy data in vascular modalities . . 40

5.8 Minutiae pre-selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405.9 Chapter Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

6 Bloom filter-based indexing 446.1 Bloom Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446.2 System Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

6.2.1 Template Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . 466.2.2 Tree Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476.2.3 Tree Traversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486.2.4 Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

6.3 State-Of-The-Art Bloom filter approach . . . . . . . . . . . . . . . . . . . . . . 506.3.1 Tree selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516.3.2 Quick traversal direction decision . . . . . . . . . . . . . . . . . . . . . . 52

6.4 Tree root emulation and random template emulation . . . . . . . . . . . . . . . 536.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

7 Spectral Minutiae CPCA-Tree indexing 567.1 System Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

7.1.1 Binary SMR-CPCA merging . . . . . . . . . . . . . . . . . . . . . . . . . 577.1.2 Masking out most common bits . . . . . . . . . . . . . . . . . . . . . . . 587.1.3 Template Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . 597.1.4 Tree Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597.1.5 Tree Traversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

viii

7.1.6 Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607.2 Template protection properties . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

7.2.1 Unlinkability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 617.2.2 Renewability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 647.2.3 Irreversibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

7.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

8 Experimental Setup 688.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 688.2 Palm Vein Feature Detection Pipeline . . . . . . . . . . . . . . . . . . . . . . . . 708.3 Excluded Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

8.3.1 Capturing Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 708.3.2 Preprocessing Failures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

8.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 728.4.1 Conducted experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 728.4.2 Experiment vocabulary . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

8.5 Workload metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 758.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

9 Results 779.1 Baseline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

9.1.1 Score Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 779.1.2 EER and ROC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

9.2 Spectral Minutiae Representation . . . . . . . . . . . . . . . . . . . . . . . . . . 809.2.1 PolyU dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 819.2.2 CASIA dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 839.2.3 PUT dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 849.2.4 Aging / Further experiments . . . . . . . . . . . . . . . . . . . . . . . . 87

9.3 Workload reduction by feature reduction . . . . . . . . . . . . . . . . . . . . . . 889.3.1 Workload baseline for the PSML-pipeline . . . . . . . . . . . . . . . . . 889.3.2 CPCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 899.3.3 Binary SMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 909.3.4 Binary SMR-CPCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 919.3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

9.4 Bloom filter-indexing Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 929.4.1 Binary SMR-CPCA Bloom filter-indexing . . . . . . . . . . . . . . . . . 939.4.2 Binary SMR Bloom filter-indexing . . . . . . . . . . . . . . . . . . . . . 989.4.3 Binary SMR-CPCA Bloom filter-verification . . . . . . . . . . . . . . . . 1029.4.4 Binary SMR Bloom filter-verification . . . . . . . . . . . . . . . . . . . . 104

9.5 CPCA-Tree-indexing Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

ix

9.5.1 Basic implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1059.5.2 Tree pre-selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1069.5.3 Additional binary comparison . . . . . . . . . . . . . . . . . . . . . . . . 1089.5.4 Masking out most common bits . . . . . . . . . . . . . . . . . . . . . . . 1099.5.5 Additional real-valued comparison . . . . . . . . . . . . . . . . . . . . . 109

9.6 Using the full vascular pattern in the SMR . . . . . . . . . . . . . . . . . . . . . 1129.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1139.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

10 Discussion 12010.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

10.1.1 Poor performance on the CASIA and PUT datasets . . . . . . . . . . . . 12010.1.2 Towards SMR on palm-print datasets . . . . . . . . . . . . . . . . . . . . 12010.1.3 SML minutiae reliability quality data impact . . . . . . . . . . . . . . . 12110.1.4 Ageing and environmental effects in the datasets . . . . . . . . . . . . . 12210.1.5 Small number of templates per Bloom filter tree . . . . . . . . . . . . . . 12410.1.6 Results compared to state-of-the-art approaches . . . . . . . . . . . . . . 124

10.2 Workload of real-valued SMR templates . . . . . . . . . . . . . . . . . . . . . . 12610.3 Derived further research topics . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

10.3.1 (Palm) vein modalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12710.3.2 Other modalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

10.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

11 Conclusions 130

A Appendix 132A.1 Bloom filter-indexing - Quick traversal direction decision first child favouring . . 132A.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

Bibliography 172

x

Chapter 1

Introduction

The steadily growing interest in biometrics for two decades lies in its promising applicationin a vast range of disciplines. At the time of writing, biometrics are already reliably used asan alternative or extension to traditional knowledge and token-based access control systems,identity documents and forensic sciences. While the biometric market value in 2015 was estim-ated at approximately 14 billion USD [Tra14], more recent studies predict market values fromapproximately 35 billion USD by 2022 [Cre16] up to 70 billion USD by 2025 [Tra17]. One ofmany catalyst for the rapid market value increase is government-driven, large-scale biomtericdeployments like the Indian AADHAAR project [Ind], which aims to enroll the entire Indianpopulation of 1.3 billion individuals and already has enrolled over 1 billion subjects, as wellas several immigration programmes. The operation of such large deployments yields immensecomputational load. Up-scaling the hardware in terms of computing power quickly reaches cer-tain limits. Therefore, the underlying systems software needs to implement efficient strategiesto reduce its computational load. Traditional indexing or classification solutions are ill-suited:the fuzziness of the biometric data does not allow for naïve hashing or comparison methods.This matter is the key motivation and the main focus of this thesis.One emerging biometric characteristic, that steadily increases market share1 and popularityis the vascular (blood vessels) pattern in several human body parts. Wrist and back-of-handvessels hold the most interest, since they are intuitive to use for users and feature several ad-vantageous properties. While several biometric modalities like the fingerprint [CFM11] and iris[DRB17] are already covered by workload reduction research, the (palm-) vein characteristiclacks approaches for such computational load reduction. Wrist, finger and back-of-hand vascu-lar biometrics are already commonly covered in several research topics, whereas the palm veinis often passed or ignored and replaced by palm print biometrics. Therefore, this thesis shalladdress the palm vein characteristic with a focus on the biometric identification scenario.

12014[Tra14], 2016 [Cre16], 2017[Tra17]

1

1.1 Thesis ContributionThis thesis contributes to current palm vein and minutiae-based biometric research by thefollowing means:

• A state-of-the-art survey of existing approaches within minutiae-based biometric workloadreduction.

• A proposal of a palm vein signal processing subsystem for minutiae-based vascular bio-metric systems based on well-known and optimized algorithms of various image processingdisciplines.

• A proposal on how to extract quality information for vascular data.

• Implementation and open-set evaluation of the proposed palm vein signal processingsubsystem.

• Re-implementation and open-set evaluation of the Bloom filter and binary tree-basedbiometric workload reduction.

• Proposal, implementation and open-set evaluation of a less complex (compared to theBloom filter-indexing approach) biometric workload reduction approach using binary treesfor all minutiae-based biometric systems.

1.2 Thesis OrganisationThis document is organised as follows:

• Chapters 2 and 3 aim to introduce the fundamentals of biometric systems with extensionto the vascular biometric characteristic and relevant related work.

• Chapter 4 proposes a palm vein signal processing subsystem capable of a robust fea-ture extraction, even in palm-print environments with a high level of noise, yielding theSpectral Minutia Representation (SMR) templates outlined in chapter 5.

• Chapters 6 and 7 describe two biometric indexing approaches used in this project.

• Chapters 8 and 9 present the experimental setup and results.

• Finally, chapter 10 concludes the thesis with a discussion of the results with proposalsfor further research topics, followed by a summary of the achievements in this project inchapter 11.

2

Chapter 2

Biometric System Fundamentals

This chapter provides a general introduction to biometric systems with a focus on the vascularnetwork (rete venosum), especially the vascular network of palm veins (rete venosum dorsalpalma manus) as the biometric characteristic of choice.

2.1 Veins as a Biometric CharacteristicThe rete venosum is the network of (connected) veins internal to the individual’s body. Theirmain responsibility is to transfer deoxygenated blood from the tissues back to the heart. Bycontrast, arteries carry oxygenated blood away from the heart to the tissues. Figure 2.1 showsthe typical layout and layers of veins and arteries below the skin (epidermis) with labels on thedifferent types.

Figure 2.1: Distribution of the vessels in the skin (source: [GL18], figure 942).

3

(a) Office environment (b) Outdoor environment

Figure 2.2: Example FIR images of the back of hand in different environments (taken from [WL06]).

(a) FIR (b) NIR

Figure 2.3: Palm images using NIR and FIR imaging ((a) taken from [WL06]; (b) from [cas]).

The idea of using blood vessels for biometric authentication dates back when Joseph Ricesubmitted his patent for vascular technology authentication in September 1985 [Ric85]. Overtime, many approaches in vascular imaging and authentication algorithms have been intro-duced, mainly using Far-Infrared (FIR) imaging in wavelengths 50 000 nm to 1 000 000 nm orNear-Infrared (NIR) imaging in wavelength 750 nm and 2000 nm.FIR imaging of blood vessels uses heat radiation emitted from the vessels and thus it does notneed external illumination. A downside of the FIR imaging approach is the high environmentalimpact, as shown in figure 2.2, which yields a very unstable image of the vascular patterndue to strongly varying contrast [Har12]. Furthermore, using FIR for palm vein imaging canlead to several problems, i.e. different skin thickness that result in different contrast levels ina single image, renders FIR imaging a much less trivial approach compared to NIR (comparefigures 2.3a and 2.3b) imaging due to different heat levels. Therefore, this thesis will focus onapproaches and algorithms using NIR images.Current NIR vessels imaging devices capture the superficial blood vessels using an NIR-enabled

4

(a) 840 nm (b) 880 nm

(c) 940 nm (d) 940 nm

Figure 2.4: Example images from palms captured with some NIR enabled devices with illuminationaround 840 nm to 940 nm.

capturing device (e.g. CCD/CMOS sensor without an IR filter) while illuminating the Regionof Interest (ROI) with light in a spectral band between 750 nm and 950 nm. These capturingapproaches base on Lunnen’s advice for medical photographers [Lun61]. He also mentions theabsorption of various NIR wavelenghts by blood based on the level of oxygen in the haemo-globin. Due to their size, location and oxygen level, the veins of the superficial blood vesselsare visible using the mentioned approach. These veins are visible as blurred1 dark lines in theimages.

To look upon the feasibility of using blood vessels, especially the palm veins, as a biometriccharacteristic the properties proposed in [BP98] will be analysed.

Universality - Is every individual able to present the biometric characteristic? Excludingcongenital malformations, diseases or amputations, every individual has hands that con-tain blood vessels. There are no relevant statistics stating the percentage of individualsin ownership of at least one hand supplied with blood in the entire human population.Nevertheless, it is safe to consider the palm veins as a universal biometric characteristic.

1The NIR radiation scatters while penetrating deeper into the skin, which leads to blurred edges.

5

Distinctiveness - Is it likely, that two individuals are not distinguishable? Even though thedevelopment of blood vessels is not truly random, the difference between individuals isgreat. There are no observable identical patterns between siblings, ancestors or even theleft and right hand of one individual [Nad07, Bad06]. In all three databases used in thisthesis (see section 8.1) there were also not any similarities found. Therefore, it is a validassumption that palm veins are a distinctive biometric characteristic.

Performance - How accurate is the biometric identification? Watanabe et al. (Fujitsu) statea False Acceptance Rate (FAR) of 0.01% and False Rejection Rate (FRR) of 0.00008%(at maximum) in [WESS05] for their PalmSecure biometric system in a verification scen-ario. However, these reports cannot be verified: the database of 140 000 palms and theirapproaches are not public. Further, a more transparent evaluation of the Fujitsu PalmSe-cure biometric system in [The06] reports a False Non Match Rate (FNMR) of 4.23% at≈ 0.01% False Match Rate (FMR). Other vascular modalities perform at a similar level.In [Nad07], Nadort presents an overview of several vascular modalities and their perform-ance stated as FAR and FRR. Summarising, vascular biometrics perform at a very highlevel. Palm vein biometrics should be no exception.

Permanence - Does the characteristic alter in time? Like other body parts, the palm veinpattern stops developing, and thus changing, on average by the age of 20. Having impliedno diseases, wounds or surgical interventions alter the vascular pattern, the only expectedchanges are shrinking and expansion of the vessels due to temperature and blood pressure.However, several publications report damages of the vascular cells of smoking individuals[Pit00, PGS+09].

Collectability - Is it easy to obtain a sample of the biometric characteristic? Acquiring a vas-cular image is fairly simple with cheap components using NIR-Imaging. However, thereare several pitfalls in acquiring a high-quality vascular image for a high performance bio-metric system. A brief elaboration of recommendations for acquiring high-quality vascularimages is outlined in section 4.1.

Acceptability - Do users accept the biometric system? The rising acceptance and trust in bio-metric systems also benefits vascular biometric systems, even thought vascular biometricsystems are not as well-known as iris or fingerprint systems. Vascular imaging can bemade contact-less, and thus hygienic, which also helps the acceptance. Several case stud-ies by Fujitsu Limited report a high acceptance in high-security applications like ATMs[csp13b] and access control [csp13a, csp14].

Circumvention - How hard is it for a attacker to acquire a sample of the biometric character-istic or to get accepted by the biometric system? Allthought it is very difficult to receivea high-quality vascular image of an individual without their knowledge, it is still possible.

6

With some knowledge, it is possible to produce artefacts (i.e. persentation attack instru-ments) from low-quality vascular images or random artefacts for presentation attacks.Countermeasures for presentation attacks are liveness detection with e.g. the pulse.

Collectability and acceptability are two important properties, why the palm veins are of highinterest. Other modalities using vascular patterns as characteristic show several drawbacks inthose properties. One example of such modality are the facial veins. The facial veins are thinner,therefore require a image capturing system with a higher resolution to detect these thin veins.Further a individual may rather lift his hand in front of a sensor then his face.

Summarized, the palm vein is a promising biometric characteristic that mostly suffers fromits lack of widespread deployment and lack of publicy-available high-quality databases (seesection 4.1 and 8.1) for research tasks.

2.2 Generic Biometric SystemThe following subsections introduce a generic biometric system and its basic operational details,which can be used to describe a vast majority of biometric systems.

2.2.1 Workflow

ISO/IEC describe a basic workflow for a generalised biometric system in their standard onbiometric performance testing and reporting [iso06]. This workflow, as shown in figure 2.5, canbe applied to every biometric modality.

Figure 2.5: Generic biometric system (source: [iso06])

The five biometric subsystems described by ISO/IEC are briefly outlined below regardingto the palm veins as biometric characteristic.

7

Data Capture Subsystem An NIR-enabled imaging device (typically monochromatic), vari-ous NIR light sources and its hardware to capture the necessary information needed forthe following subsystems (see section 2.1).

Signal Processing Subsystem A pipeline of algorithms is needed to extract the biometricdata (feature set) and process it to an understandable model, the template, for the sub-sequent subsystems. This subsystem can further be detailed to the following processingsteps:

Segmentation/ROI detection The image retrieved from the Data Capture Subsystemhas to be segmented to the foreground and background. Subsequently, the ROI hasto be found. This step is crucial for further processing steps, since it is necessary toalways process the nearly exact portion of the palm to extract a stable biometricfeature set.

Feature extraction After the ROI has been successfully found, the biometric featureset has to be extracted. For vascular biometric systems, this can be e.g. the wholevein pattern or specific features like bifurcations or endpoins, so-called minutiae, ofthe pattern. The requirement for the biometric feature set is to have a low intra-classvariation (low variance in samples from the same individual) while maintaining ahigh inter-class variation (high variance in samples from different individuals).

Quality Control It is possible that the feature extraction fails to extract sufficient fea-tures due to poor image quality or a misslocated ROI. Automated or manual qualityassessment can reject low-quality feature sets to maintain an overall good perform-ance of the biometric system. In vascular biometric systems, it is possible to countthe number of veins or minutiae found, or in an early step check for sufficient contrastin the ROI to reliably extract the vein pattern.

At the end of this pipeline, the Signal Processing subsystem returns a template createdfrom the extracted biometric feature set. The template needs to address various privacyconcerns, e.g. it should not be possible to translate the template back to an image of theoriginal vessels [FZ05].

Data Storage Subsystem Once the data is processed into a template, it has to be stored sothat the comparison subsystem can access it on demand. The Data Storage Subsystemis subject to several privacy concearns outlined in section 2.2.4.

Comparison Subsystem If a biometric system is queried with a request, it has to comparethe biometric probe processed by the signal processing subsystem with the templatesstored by the data storage subsystem. Various algorithms have emerged over time usingall kind of different approaches and trade-offs. Most algorithms result in a probability,rated by a similarity or dissimilarity score.

8

Decision Subsystem Once all candidate templates are compared with the biometric probe,the decision subsystem has to evaluate whether a template and the probe can be matedand if there is a match in the candidate list or the biometric probe is non-matched.

The workflow also describes three paths in the biometric system: (i) Enrolment; (ii) Veri-fication; and (iii) Identification;

Path (i) describes the process of creating a reference template, will be stored in the enrolmentdatabase. This reference template is stored inside the Data Storage Subsystem and later on usedas a reference in comparison to a query. Paths (ii) and (iii) are subject to section 2.2.3.

2.2.2 Subsystems in this thesis

Through this thesis, implementations and approaches for the different subsystems described in2.2.1, targeting the palm veins, are presented. To help the reader’s overview of the differentsubsystems, the graph in figure 2.6 is used. Every node in this graph is expanded in its corres-

Data Capture Subsystem

Signal Processing Subsystem

Data Storage Subsystem

Comparison Subsystem

Decision Subsystem

Figure 2.6: Collapsed subsystem-graph, based on [Har12].

ponding chapter to guide through the details (in the further course referenced as a pipeline). Aconcrete introduction in the Data Capture Subsystem is beyond the scope of this thesis and willbe skipped. Refer to section 2.1 and 4.1 for a brief overview or [Har12] for a deeper introductionof the Data Capture process in (palm-)vein modalities.

2.2.3 Operation Modes

As already introduced in section 2.2, two more paths through the biometric system exist. Thetwo leftover paths (ii) and (iii) are called operation modes. They describe the flow, selectionand processing of information in the system and determine how the Decision Subsystem takes

9

a decision. In an abstract sense, they describe the difference between an 1 : 1 (O(1)) and 1 : S(O(S)) biometric system.

Verification In verification mode, a biometric system confirms the identity claim of an in-dividual. The individual provides a claim to an identity, e.g. with an username, RFIDcard or another attribute only assigned to him. With this claim, the biometric systemonly needs to compare the individual’s query with the reference templates stored for theclaimed identity. For palm vein biometric systems, this corresponds to an O(1) (only onepalm enrolled) or O(2) (two palms enrolled per individual) comparison.

Identification In identification mode, a biometric system has no identity claim by an indi-vidual. The individual only presents his biometric characteristic and queries the biometricsystem with this probe. Now the biometric system has to confirm that the individual isenrolled and find his identity or deny the probe if the individual is not enrolled in thebiometric system. For this purpose, in a naïve approach the biometric system has to com-pare the query template with all reference templates of S individuals. In the worst case,O(S) (only one palm enrolled) or O(2S) (two palms enrolled per individual) comparisonsare needed to reach a decision.

Without rotation compensation (like pre-alignment of the probe or rotation invariant rep-resentations) of the biometric probes and references, the complexity rises by a factor of thenumber of rotations that have to be tested for both operation modes. While computationalcost is a non-issue for biometric systems in verification mode, these costs are intolerable inlarge-scale identification mode biometric systems. The naïve approach also introduces anotherissue presented by [Dau00]. Daugman demonstrates the probability PS of having at least oneFalse Match (FM) in an identification scenario with equation (2.1). S denotes the number ofenrolled subjects and P1 the probability of a false match in a single identification transaction.

PS = 1− (1− P1)S (2.1)

Figure 2.7 shows the plot for equation (2.1) with a FMR of 1%, 0.1% and 0.01%.Looking upon a biometric system with S = 800 enrollees, the probability of at least one

false match at a FMR of 1% is P800 ≈ 99.9%. Even with a FMR of 0.1% the probability is stillvery high (P800 ≈ 55%).

It is clearly visible, that the probability of at least one false match quickly rises with thenumber of enrolled subjects when choosing the nïve approach. Without an reduction of tem-plate comparisons, large biometric systems quickly reach a point where they won’t behave likeexpected: the system could fail to identify the correct user or, even worse, allows access to anunauthorized individual. While this is less an issue for very small biometric systems (S < 10),larger systems need to reduce the number of template comparisons in a identification or Du-plicate Enrolment Check (DEC) scenario to tackle computational workload and false positive

10

100 101 102 103 104 1050

0.2

0.4

0.6

0.8

1

S

PS

P1 = 1%P1 = 0.1%P1 = 0.01%

Figure 2.7: Plot of equation (2.1).

occurrences.

2.2.4 Template protection requirements

Biometric templates carry sensitive information from individuals, and thus they are subject toseveral privacy concerns, including the following:

Impersonation Biometric templates can be stolen or forged like passwords, thus enabling anattacker to generate an artefact (i.e. replica of the biometric characteristic) and severalattack scenarios from a template.

Irrevocably Biometric characteristics cannot be changed.

Tracking With a stolen template, a attacker can track people’s behaviour by cross-comparisonof templates across several databases.

The ISO/IEC 24745:2011 (draft: [Bus10]) describes several requirements to protect templatesto prevent these privacy concerns.

Renewability In the event of a compromised template, it should be possible to generate anew and different template from the same biometric characteristic.

Unlinkability The templates stored in different biometric databases and applications shouldnot be linkable, e.g. to prevent tracking.

Irreversibility A template needs to be transformed by an asymmetric transformation in a dif-ferent and comparable representation to prevent a reconstruction of the original biometricsample from a template while maintaining the ability to compare different templates.

Refer to [RU11] for a overview of some existing template protection schemes.

11

2.3 Chapter ConclusionThe vascular pattern of the humans palm is a promising biometric characteristic in the meansof the seven properties universality, distinctiveness, performance, permanence, collectability,acceptability and circumvention. Biometric probes are collectable with NIR imaging when illu-minating the palm in wavelengths of 750 nm to 950 nm. The vascular network appears as darklines in NIR imaging.

Every biometric system can be generalised to the five subsystems data capturing, signal pro-cessing, data storage, comparison subsystem and decision subsystem operating in verificationor identification mode. A biometric system has to address several privacy concerns; templatesshouldn’t be traceable across different databases, templates shouldn’t be reversible to its bio-metric sample and it should be possible to renew templates in a way, that a renewed templateisn’t linkable to a old template of the same instance.

Section 2.2.3 introduced the at-least-one-false-match problem in identification scenarios,which also applies to the palm vein modality. Thus, it is required to employ a strategy to reducethe number of necessary template comparisons (workload reduction) for the palm vein modality.Workload reduction for palm vein modalities remains an insufficiently-researched topic. In thefollowing chapter, selected state-of-the-art approaches for workload reduction and state-of-the-art representations of (palm) vein patterns will be presented, combined and expanded to findnew strategies for workload reduction in palm vein environments.

12

Chapter 3

Related Work

This chapter describes current state-of-the-art and newly -introduced workload reduction ap-proaches in palm vein identification scenarios, preceded by a brief discourse in palm vein featurerepresentations. The chapter is organized in three sections: first, a generalized list of state-of-the-art types of palm vein representations is presented in section 3.1 and discussed. Second, abrief overview of types of workload reduction is presented in section 3.2, followed by a shortliterature survey for selected types in the corresponding subsections.

3.1 Vascular Biometric SystemsAt the time of writing, there are two common features found in palm vein patterns.

Minutiae Points Vein minutiae points are nearly identical to fingerprint minutiae points:they mark endpoints, bifurcations and trifurcations of the vein patterns. Minutiae pointscan be described by their location and rotation.

Vein Strands The veins themselves can be used as features. Descriptors to veins are start-and endpoints, direction, curvature, angles or abstract figure.

Due to their almost identical characteristics to fingerprint minutiae, the vein minutiae arecommonly used for vascular biometric systems. Substantial research effort has been devotedto fingerprint minutiae and a broad segment of the introduced approaches are applicable, butnot necessarily feasible, to palm vein minutiae: recalling section 2.1, compared to fingerprintminutiae, the palm vein minutiae suffers from a higher noise in the capturing process due to theblurred edges of the veins and low contrast. Therefore, most fingerprint minutiae approachesneed to be re-evaluated for the more fuzzy palm vein minutiae. A brief survey of differentfingerprint minutiae approaches is presented in [Zae11].Using minutiae as feature descriptor isn’t as commonly used as using the full vascular patternas feature descriptor. For example, patents [Wat14, JHF14, SKS+14] by Fujitsu Limited hintat the usage of the full palm vein vascular pattern as a biometric feature set in their comercialpalm vein biometric system.

13

3.2 Workload ReductionAs already mentioned in 2.3, it is crucial to reduce the number of necessary template compar-isons in a biometric system. The different approaches in workload reduction are categorized inthree different types by [DRB17].

Serial combination of algorithms Instead of comparing every template with the query us-ing the selected comparison algorithm, a less complex, respectively faster, and approxim-ate algorithm is used to find a shortlist of most likely template matches. subsequently,only the templates in the shortlist are compared using a complex and more accuratealgorithm.

Classification, Clustering and Binning Another two-step approach is to split the enrolleedatabase into multiple subsets based on a easy-to-determine feature characteristic. On aquery, the probe is classified and only the enrolled templates matching the class of theprobe are compared with the probe using the actual comparison algorithm. Trivial classdescriptors in palm vein templates are e.g. the race or left/right hand.

Indexing Workload reduction utilizing probabilistic or hierarchical data structures (e.g. trees).These approaches reduce the system load in terms of the big-O notation.

It has to be noted that even if the serial combination of algorithms and classification typesappear to be the same (both types run a pre-selection step to select a shortlist of qualified tem-plates), there is one crucial difference: while the serial combination of algorithms generates atailored shortlist for each query from the entire database, the classification/binning approachespredefine shortlists of the database and select the right list for a query.

For the completeness sake, [DRB17] also lists hardware acceleration and parallelism as acategory: disjoint parts of the enrollee database can be processed simultaneously on multipleCPUs/threads by using multi-core processors or GPUs. However, it is noted that the workloadis not reduced but simply distributed so the problem of reducing the number of templates isnot solved by these approaches. Further, [DRB17] divides the approaches in image encodingdepended and independent approaches.Biometric systems can use a combination of the presented categories. For example, precedinga paralleled indexing approach with an classification approach may further reduce workload inlarge-scale scenarios.Although it appears obvious, it has to be noted that a workload reduction approach must not(significantly) impair the naïve biometric performance to be viable.The following subsections provide a short list of examples for the different types of workloadreduction and list some possible advantages and disadvantages. Note that most examples arefor other modalities that use either veins as a biometric characteristic or use minutiae-points as

14

feature descriptor, since approaches explicitly targeting palm veins are not broadly researchedor publicly available.

3.2.1 Serial combination of algorithms

All surveyed approaches that are stated as a serial combination of algorithms by the authors arein fact indexing or classification approaches, followed by a state-of-the-art biometric recognitionapproach.

3.2.2 Classification, Clustering and Binning

The advantage of the classification, binning and clustering approaches compared to the serialcombination of algorithms is trivial: while a serial combination of algorithms has to inspectthe whole database just-in-time, the predefined classes of a classification/clustering or the com-puted bins of a binning approach are generated off-line during the database construction. Thisleads to increased computational overhead while creating the database, although the overheadfor each query is comparatively small. Frequently, classification and clustering are used assynonyms for each other. However, in data mining, classification (supervised) and clustering(unsupervised) are two different learning approaches. Classification approaches are trained withdetermined outcomes (i.e. pre-defined classes), while in clustering the algorithm has to determ-ine the classes by itself.

A fairly new approach in workload reduction in terms of binning is presented by [ZLF+14].By extracting an orientation matrix based on the Gaussian Radon transform using the Mod-ified Finite Radon Transform (MFRAT) [JHZ08, ZK11], the principal direction features of apalm vein image is determined. Based on the angle of the principal direction, the template isclassified in six bins (Binα), corresponding to the major directions 0°, 30°, 60°, 90°, 120° and150°. Therefore, a probe that is classified as Bin30° only needs to be compared to the referencesthat are classified as Bin30°.Another classification approach presented in [SRB15] uses clustering (K-Means [Llo82]) to formgroups of similar finger-vein patterns. Thus a query first is compared to the centre points of eachcluster and second the query is only compared to the templates stored in the most fitting cluster.

Disadvantages of a classification workload reduction approach include

the limited number of classes for each class descriptor: differing between left or right hand onlyresults in two classes; using the skin colour may result in two to four classes, wherebyeven the introduced approach using the principal direction of the palm veins only featuresfive classes. Thus, the expected workload reduction for classification is comparativelysmall. Furthermore, most classification approaches expect a evenly distributed number of

15

templates in every class. In most real-world scenarios, this is not the case, and thus theactual workload is much higher when one query is classified to a large bin.

the similar two-step problem like in serial combination of algorithms approaches: if a shortlist does not contain the correct reference template, the following steps would yield anon-match, thus lowering the genuine acceptance rate.

3.2.3 Indexing

Indexing on biometric data requires different approaches than common indexing algorithmsdue to the fuzziness: instead of querying with the same key, a similar key is queried and thusnormal hashing approaches are not feasible. The retrieval performance of an indexing schememost depends on its hashing functions, which need to fulfil the conditions that the hashedrepresentation of two similar templates are also similar as well without losing too much inter-class variance. A good introduction1 and a possible approach to the problem is presented by[HDZ08]. Most noteworthy disadvantages and trade-offs are the high offline (preprocessing) andadditional storage costs.

An novel minutiae-based indexing approach is presented in [CFM11] using the MinutiaCylinder-Code (MCC) (presented in [CFM10]) minutiae representation. Initially targeting fin-gerprint minutiae, the design of the MCC offers promising characteristics for palm vein minu-tiae. The MCC is a fixed-radius local minutiae approach represented in a 3D cylindrical datastructure. These cylinders are built from invariant distances and angles in a neighbourhood ofeach minutiae, without the type and quality.

The authors state, amongst others, three advantages that - projected to palm vein minutiae- deal with some of the negative characteristics of palm vein feature detection:

• The fixed-radius approach tolerates missing and spurious minutiae.

• Small feature extraction errors are tolerated due to the adoption of smoothing.

• Noisy regions with many spurious minutiae are dealt with a limiting function.

Indexing the MCC is achieved by linearising the cells of a cylinder corresponding to a givenminutiae, thus receiving a vector of linearised cylinders. Each linearised cylinder is hashed bythree different hashing functions. For each hash, the value (bucket) is stored. To generate ashort list of candidate templates, the same procedure is applied to the query and the databasetemplates with the highest number of bit collisions are selected.

1Another good publication for the awareness of the fuzzy data problem can be found in [Pro13].

16

3.3 Chapter conclusionsThis chapter has introduced the different types of workload reduction and presented some cur-rent state-of-the-art biometric workload reduction approaches for minutiae- and vascular-basedmodalities. The three introduced types are serial combination of algorithms, classification/bin-ning and indexing. For each presented approach, the basics were briefly outlined. Some presen-ted approaches base on special representations. This is most likely found for indexing and serialcombinations of algorithm types. Most classification and binning approaches do not rely ona special representation, and thus can be efficiently employed without additional storage re-quirements. This renders them a good choice as a first level for multi-level workload reductionenvironments.

Article Type stated Real type Performance Workload Non-intrusive*

[CFM11] Indexing Indexing — — No[ZLF+14] — Classification — 16.6% Yes[SRB15] — Classification (Clustering) — — Yes* Does not requires a special representation, thus easily combinable with other approaches.

Table 3.1: Summary of surveyed workload reduction approaches.

It has to be noted that many approaches labelled as a serial combination of algorithmsturned out to be binning approaches, hence the relabelling in table 3.1.Before presenting the two indexing workload reduction approaches used for the practical workof this thesis, the feature extraction pipeline and the feature representation will be presentedin the next two chapters.

17

Chapter 4

Palm Vein Feature Detection

As shown in figure 4.1, the pipeline of a common Signal Processing Subsystem can be furtherseparated into four different stages: quality control, pre-processing, feature extraction and post-processing. All approaches in this chapter belong to the pre-processing or feature extractionstage. The Signal Processing Subsystem in palm vein modalities is much more complex and

Data Capture Subsystem

Signal Processing Subsystem

Data Storage Subsystem

Comparison Subsystem

Decision Subsystem

Quality control

Pre-Processing

Feature Extraction

Post-Processing

Figure 4.1: Signal Processing Subsystem

error-prone compared to other biometric modalities. Even compared to other vein modalities,there are some additional difficulties to address. Compare figures 4.2a and 4.2b. Note the in-homogeneous illumination of the ROI due to the geometry of the palm. Moreover, note thecreases and strong skin valleys resulting in large amount of noise that can not be found in theskin texture of the wrist or the back of the hand. The lack of public available large-scale, high-quality palm vein databases requires to use databases built for other modalities like palm print(refer to section 8.1 for details) in experiments. Palm print databases purposely feature a highamount of skin details, and thus feature more noise than purpose-built palm vein databases,which aggravates the palm vein feature extraction (recall, figure 4.3a sample of a purpose-build

18

(a) (b)

Figure 4.2: Raw samples of different vein-modalities found in public databases: (a) wrist [PUASRL10]and (b) palm [cas].

palm vein imaging device; figure 4.3b sample of a purpose-build palm-print imaging device).A discussion of different approaches for low-quality vein images in [KK10] presupposes highcontrast and little, homogeneous noise, and thus is not feasible for the fuzzy-noise environ-ment of palm-print images. Further related work by [HOXB11, HOX+12] has already brieflydiscussed several approaches for pre-processing and feature extraction of other vein modalitieswith less noise. However, the lack of documentation details on these approaches exacerbatesthe reimplementation and replicability for further research.Since the pre-processing and feature extraction steps are not trivial for a high-noise environ-ment and important for the biometric performance of following subsystems, and given thatmost approaches are taken from other modalities, it is necessary to discuss the adaptationsand the combination of these approaches in the following sections. The first following sectionprovides a brief elaboration of recommendations for NIR imaging to increase the overall inputquality.

4.1 Recommendations for high-quality vascular imagingFor high-quality images of the superficial blood vessels, a NIR-pass filter that blocks visiblelight is recommended to filter noise like reflections and skin details. Another recommendationfor noiseless images is to illuminate the ROI with a diffuse, rather than specular, illuminationsource. Further, a very small back-focus (in respect to the skin surface) of the lens helps tofurther reduce noise. Figure 4.3 shows the difference in cleanliness of the captured veins whenusing diffuse (4.3a) and specular (4.3b) sources.

Observe the noise in figure 4.3b: The image features parts of the skin texture (mostlyshadows in the texture valleys) and suffers from inhomogeneous illumination distribution dueto specular light, which can lead to false detection of veins by the feature detection subsystem.

19

(a) (b)

Figure 4.3: Difference between diffuse and specular lightning of the palm ROI using reflection captur-ing: (a) diffuse NIR illumination (captured with a proprietary device); (b) specular NIR illumination(from [cas])

4.2 ROI DetectionIn real-world contact- or guidance-less palm vein applications, a robust and accurate ROI de-termination is a crucial step in the pre-processing stage.The single requirement of the introduced approach to the user is that he should spread1 hisfingers. It can safely presupposed that in high-security applications, individuals keep a highlevel of cooperation and are aware of requirements in interacting with the biometric system.Starting with a prioritized watershed-flood filling - described in [BLM14] - for image segment-ation, the palm and fingers are separated from the background and other noise using a simplemask shown in figure 4.4. Once the palm and fingers are filled (visualized in 4.5adg), the Border

Figure 4.4: Mask used for watershedding, rectified (white = highest priority; black = lowest priority).

Following algorithm by Suzuki from [S+85] is used to retrieve the contour of the hand. Next, theconvex hull of the hand is retrieved using the approach from [Skl82]2. With the convex hull and

1Overspreading should be avoided to keep the palm geometry undistorted as possible.2Note that [GY83] describes a more efficient implementation for finding the convex hull that could be used

to replace the algorithm used.

20

the real contour of the hand, so-called convexity defects can be calculated. Convexity defectsare gaps, thus the error, between the convex hull and the contour. They can now be used tofind the finger gaps. In figure 4.5, the convex hull (purple) and the convexity defects (green)are visualized for three different samples in the second column. Finding the three finger gapsin a vector of convexity defects can be implemented according to the following assumptions:(i) the convexity defects of fingers are comparatively deep (large error); (ii) the convexity de-fects of interest have a narrow start and end point (start/end of error); (iii) convexity defectsof fingers lay very close together; (iv) the center convexity defect (between the middle andring finger) has the smallest distance to the other two finger convexity defects; and (v) whenspanning a line between the two outer gaps, the middle gap has to be above the line, and thusthe orientation of the hand can be determined. If the assumptions are implemented in thisorder, the three finger gaps found can easily be mapped to the left, middle and right fingergap. The reference points (shown coloured blue and cyan in the third column of figure 4.5)can now be used to span the ROI-rectangle: left and right mark the ROI’s size with rotationand the middle reference point helps to decide where to move the rectangle. Now the ROI-center can be calculated: Let Pcenter = (Pleft + Pright) ∗ 0.5, ∆P the distance between Pleft andPright, and α the angle of the axis spanned by Pleft and Pright in respect to the X-axis, thenPROIxy = (Pcenterx + ∆P ∗ 0.8 ∗ cos(α), Pcentery + ∆P ∗ 0.8 ∗ sin(α)). The scalar 0.8 is used toreduce the translation, otherwise the ROI-center would translate too far away from the centerof the palm. Around Pcenter, a rectangle with the edge size of 120% of ∆P is created and rotatedby α. These calculations result in the blue rectangles shown in the third column of figure 4.5.Note that there are two blue rectangles per palm: the enclosed area between the two rectanglesis used to determine whether the calculated ROI is placed too close to the palms boundaries.

The introduced ROI determination approach proved robust and accurate on multiple samplesand correctly framed3 the ROI of 99.2% (∼ 1190/1200) of the 940 nm samples contained in theCASIA-MS[cas] database. Another feature is its rotation invariance: the ROI is correctly de-termined independently of the image and the palm rotation.

4.3 Image EnhancementThe second stage in the pre-processing step involves various image enhancement operations.As already mentioned in the fundamentals chapter (see section 2.1), vascular imaging is verynoisy and blurred. To best eliminate noise and receive a most clean and stable vascular image,the following combination of the Non-Local Means (NL-means) and the Non-Linear Diffusion(NL-diffusion) algorithms are used in the Signal Processing Subsystem.

3ROI centered in the palm, no background inside the ROI and at least 90% overlap of ground-truth anddetermined ROI.

21

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

Figure 4.5: Visualisation of different intermediate data in the ROI finding process for samples fromCASIA-MS[cas]: (a,d,g) hand seperated from background using priority-watershedding with maskshown in figure 4.4; (b,e,h) convex hull (purple) and convexity defects (green); (c,f,i) original imagewith projected ROI-rectangle (blue) and reference points (green, cyan, grey).

4.3.1 Non-Local Means

The first used algorithm is used to create an illumination invariant texture of the ROI by non-local smoothing. Initially presented for face recognition by [ŠP09], it was successfully used by[HOXB11, HOX+12] for vascular imaging. For ROI-sizes of about 128px, a template windowsize of 7×7 and a search window size of 21×21 proved feasible in all experiments carried out inthe curse of this project.Figure 4.6b shows the output of the NL-means when a worst-case sample in terms of illumin-ation variance shown in figure 4.6a is used as a input. Note the shadows in the input sampleresulting from inhomogeneous illumination and a folded palm that are successfully removed.

22

(a) (b) (c)

(d) (e) (f)

Figure 4.6: Output from single stages of the image enhancement process: (a) raw input (ROI fromfigure 4.5f, see section 4.2); (d) raw input (NIR from [pol]); (b,e) output of NL-means (inverted); (c,f)denoised image output of NL-diffusion after inversion.

Another property of the presented algorithm used is its ability to increase contrast on edgeswhile removing illumination contrast, with a high decay parameter h. The decay parameter isused to control the decay of the Gaussian weighted Euclidian distance that serves as a similarityvalue between local neighbourhoods. Refer to section 3.1 of [ŠP09] for a in-depth descriptionof the decay parameter.The introduced contrast enhancement by this algorithm with h = 30 is sufficient for furthersteps. Thus, an additional contrast enhancement is not necessary. Furthermore, well-knowncontrast enhancement algorithms like CLAHE tend to increase contrast in non-vein areas, thusagain introducing additional noise that can lead to falsely-detected veins by the following steps.

A disadvantage of this algorithm is its high computational complexity. However, the ad-vantage of skipping further contrast enhancement with computational complex methods putsthis disadvantage in perspective.

4.3.2 Non-Linear Diffusion

Following the NL-means, the NL-diffusion agorithm is used to remove high-frequency and somemid-frequency noise introduced by the skin texture. Noise removal is a well-researched topic.The selected approach introduced by [Wei01] features two feasible advantages for this applic-ation: (i) reduces high- and mid-frequency noise and (ii) preserves blurred edges. Advantage(i) is expected for a good noise-reduction algorithm while advantage (ii) can be considered an

23

additional feature.

For demonstration purposes, a very noisy sample was chosen in figure 4.6d. Observe howNL-diffusion removed the high-frequency and some mid-frequency noise in 4.6f after applyingNL-means in figure 4.6e, while the veins are still clearly visible due to the ability to keep blurrededges. The NL-diffusion is configured with by the paramenter k, which describes the amountof diffusion iterations done.

4.4 Vein DetectionAfter completing the pre-processing stages of finding the ROI and enhancing the cut ROI, thefeatures can be extracted. Recall section 2.1, the features of interest (the veins) are the darklines. The task to solve can be simplified to line tracking. While line tracking is a well researchedtopic, due to its application in a vast number of disciplines most approaches are not feasiblefor vein extraction: many approaches require pre-defined starting points, directions or patternsand depend on high-contrast edges. Neither of these requirements are fulfilled in palm veinpatterns. With the introduced image enhancement pipeline, the contrast isn’t high enough onedges of small veins for traditional line tracking approaches to extract a stable representationof the vascular pattern. Even with a high enough contrast, gaps, noise and the determinationof starting-points are still a problem for line tracking approaches.

In [MNM04], a approach using simple line-tracking with random starting-points, trimmedfor vascular images, is presented. However, the approach fails to reliably follow veins in low con-trast areas and thick veins. The same authors present in [MNM07] a more robust vein-patternextraction approach using maximum-curvature points. Simplified, the algorithm examines theROI-image horizontally, vertically and diagonally line by line as a intensity graph4 and con-verts it in a curvature representation. The curvature representations is now searched for localmaxima. Every local maxima is scored with the probability that it is positioned on the centreof a vein, and thus a map of scores is created.

Using the enhanced image from figure 4.6c as an input for the maximum-curvature al-gorithm, the score map visualized in figure 4.7a is received. With a threshold function, thescore map can be converted to a binary representation of the palm vein pattern. Even thoughthe maximum-curvature algorithm tries to connect the vein centres inside the score map, exper-iments have shown that using a dilate-erode-dilate step helps to connect single gaps or closingsmall holes (see figure 4.7b).

4The very basic idea of [MNM07] is again picked up (three years later) by [KK10]. Neither [KK10] nor thefollowing publication [KK11b] mention the approach by [MNM07].

24

(a) (b) (c) (d)

Figure 4.7: Output from single stages of the feature detection process with input from figure 4.6c:(a) score map retrieved with the maximum-curvature algorithm; (b) binarised and dilated score map;(c) thinning with algorithm of Guo and Hall; (d) detected minutiae (bifurcations/trifurcations green,endpoints red);

The response of the maximum-curvature algorithm can be trimmed with its σ parameterthat controls the kernel size of the algorithm. Reducing σ decreases the size of the window inwhich local maxima are searched, thus increasing the sensitivity. In the original publications aσ = 8 is used. Experiments showed an increased Genuine Acceptance Rate (GAR) for all useddatabases with σ = 7.

Another application of the maximum-curvature algorithm response is extracting qualityinformation about the vascular pattern. Refer to section 5.7.1 for a detailed description of thequality extraction process.

4.5 MinutiaeWith the palm vein pattern received from binarising the maximum-curvature score map, theminutiae are easily extractable. Instead of complex pattern searching, a much faster and morerobust approach based on the skeletarised representation of the binarised score map can beused. The approach presented in [OHBL11] extracts the minutiae type and rotation in one im-age iteration with a single convolution kernel, where other approaches need multiple iterationsand kernels.

To receive the skeleton (see fig. 4.7c) of the vein pattern (see fig. 4.7b), the algorithmintroduced by [GH89] is used. The thinning algorithm was chosen because it best preserves theinitial structure[Koc13] and is highly parallelisable.

The kernel by [OHBL11] (see fig. 4.8) represents a 2D bitmask, thus applied to the binaryskeleton image, every pixel is assigned with a value describing its neighbourhood. Minutiae(skeleton endpoints, bifurcations and trifurcations) and their rotations can be determined byspecial values in the filters response: e.g. Endpoints at rotation 0°, 45°, 90°, 135°, 180°, 225°, 270°and 315° (ccw, starting top-centre) are assigned a value of 258, 257, 384, 320, 288, 272, 264 and

25

1 2 4128 256 864 32 16

Figure 4.8: 2D-Kernel used by [OHBL11] to receive minutiae types and rotation in one iteration.

260. Figure 4.7d shows the minutiae found. Bifurcations and trifurcations are marked green,while endpoints are marked red with an additional marker following the direction of the end-point.

4.6 Chapter Conclusions

(a) (b) (c) (d)

Figure 4.9: Pipeline applied on six samples of four palm instances selected from [pol] and rendered asone image. Each colour represents one sample. Colours: Red, green, blue, cyan, magenta, yellow.

In this chapter, the pipeline for the pre-processing and feature extraction stages of theSignal Processing Subsystem was presented. The presented pipeline is a robust basis for furtheralgorithms: Figure 4.9 shows the combined output of six samples from four randomly-selectedpalm instances. Note the robust vein extraction even under the noisy conditions of a palm-printdatabase. Most errors appear to be a result of different blood circulation due to heartbeat. Onlysome minor translation errors are observable.

26

Chapter 5

Spectral Minutiae Representation

Data Capture Subsystem

Signal Processing Subsystem

Data Storage Subsystem

Comparison Subsystem

Decision Subsystem

Quality control

Pre-Processing

Feature Extraction

Post-Processing

Figure 5.1: Signal Processing Subsystem

By now, the pre-processing and feature extraction stages have successfully extracted thepalm vein minutiae for further processing. The following Comparison Subsystem and DataStorage Subsystem can use the raw minutiae vector as a template. However, this introducesseveral problems, starting with privacy concerns in terms of storing the raw biometric featuresand ending with computational drawbacks due to differently-sized feature-vector1 sizes.To address these issues, a post-processing stage can convert the raw minutiae to a fixed feature-size representation that should not be reversible to the minutiae. This chapter describes theapproach that was investigated in this thesis. Readers already familiar with the SMR approachmay only read section 5.7 and skip the rest of this chapter.

1Even probes of the same subject differ in their number of features.

27

5.1 Spectral MinutiaeInspired by the Fourier-Mellin transform [CP76] used to obtain an translation, rotation andscaling invariant descriptor of an image, the SMR [XV10c] transforms a variable-sized minutiaefeature vector in a fixed-length translation, and implicit rotation and scaling invariant spectraldomain. To prevent the re-sampling and interpolation introduced by the Fourier transform andthe polar-logarithmic mapping, the authors introduce a so-called analytical representation ofthe minutiae set and an so-called analytical expression of a continuous Fourier transform thatcan be evaluated on polar-logarithmic coordinates. According to the authors, the SMR meetsthe requirements for template protection and allows faster biometric comparisons.

To represent a minutiae in its analytical form, it has to be converted into a Dirac pulse tothe spatial domain. Each Dirac pulse is described by the function mi(x, y) = ω(x−xi, y−yi), i =1, . . . , Z where (xi, yi) represents the location of the i-th minutiae in the palm vein image. Nowthe Fourier transform of mi(x, y) is given by:

F{mi(x, y)} = exp(−j(wxxi + wyyi)) (5.1)

Based on this analytical representation, the authors introduced several types of spectral rep-resentations and improvements for their initial approach. The following four sections will coverthe three main types and their sampling introduced in their corresponding publications. Sub-sequently, the introduced feature reduction approaches and the binarisation approach, followedby the comparison methods, are presented.

5.1.1 Spectral Minutiae Location Representation

The initial approach introduced in [XVK+08] is called the Spectral Minutia Location Repres-entation (SML). It only uses the minutiaes location for the spectral representation:

M(wx, wy) =∣∣∣∣∣

Z∑i=1

exp(−j(wxxi + wyyi))∣∣∣∣∣ (5.2)

To compensate small errors in the minutiae location, a Gaussian low-pass filter is introducedby the authors. A 2-D Gaussian filter in the space domain is represented by

g(x, y) = 12πσ2 exp

(−x2 + y2

2σ2

)(5.3)

and its Fourier transformG(wx, wy) = exp

(−

w2x + w2

y

2σ−2

). (5.4)

Thus, the magnitude of the smoothed SML with a fixed σ is defined as

M(wx, wy; σ2) =∣∣∣∣∣exp

(−

w2x + w2

y

2σ−2

)Z∑

i=1exp(−j(wxxi + wyyi))

∣∣∣∣∣ (5.5)

in its analytical representation. By taking the magnitude, the translation invariant spectrum isreceived.

28

5.1.2 Spectral Minutiae Orientation Representation

Additional to the location, the orientation θ of a minutiae can be incorporated as an additionalsource of information to better describe a minutiae. This second Spectral Minutiae type wasintroduced in [XVB+09]. For the Spectral Minutia Orientation Representation (SMO), to everyminutiae the function mi(x, y, θ) is assigned, such that

F{mi(x, y, θ)} = j(wxcosθi + wysinθi) ∗ exp(−j(wxxi + wyyi)). (5.6)

Applying the Gaussian smoothening and taking the magnitude yields

M(wx, wy; σ2) =∣∣∣∣∣exp

(−

w2x + w2

y

2σ−2

)Z∑

i=1j(wxcosθi + wysinθi) ∗ exp(−j(wxxi + wyyi))

∣∣∣∣∣ . (5.7)

However, the SMO did not show better results in experiments than the SML: the orientation isincluded as a derivative of the delta function, thus amplifying the noise in the higher frequencies.To compensate the higher noise, a higher σ is needed for the Gaussian kernel, therefore losingprecision.

5.1.3 Spectral Minutiae Complex Representation

To tackle the disadvantages of the SMO approach, the Spectral Minutia Complex Represent-ation (SMC) approach was introduced in [XV10b]. Instead of incorporating the orientation asderivative of the delta function, the orientation is assigned by a complex amplitude ejθi . Thus,the magnitude of the Gaussian smoothed Fourier spectrum yields

M(wx, wy; σ2) =∣∣∣∣∣exp

(−

w2x + w2

y

2σ−2

)Z∑

i=1exp(−j(wxxi + wyyi) + jθi)

∣∣∣∣∣ . (5.8)

The authors achieve an overall better performance using the SMC compared to the SML evenfor the lesser-quality dataset.

5.2 Spectral Minutiae SamplingAs already mentioned in section 5.1, the different types of the SMR can be evaluated on apolar-logarithmic grid. When sampling the SMR on a polar-logarithmic or polar-linear grid,rotation of the minutiae become horizontal circular shifts. For this purpose, sampling of thecontinuous spectra 5.5(SML) or 5.7(SMO) is proposed by the authors using M = 128 in theradial direction λ logarithmically distributed between λmin = 0.1 and λmax = 0.6. The angulardirection β for SML and SMO is proposed between β = 0 and β = π in N = 256 uniformlydistributed samples. A sampling between β = 0 and β = π is sufficient due to the symmetry ofthe Fourier transform for real-valued functions.The SMC features much more information in the higher frequencies. Instead of sampling λ ofSMC on a logarithmically distributed grid, sampling on a linear grid is proposed. This provides

29

more samples in the higher frequency part. Furthermore, due to the additional complexity, theSMC is proposed to be sampled between β = 0 and β = 2π in N = 256 uniformly distributedsamples.

(a)

(b) (c)

(d) (e)

Figure 5.2: Illustration of the Spectral minutiae approach: (a) input minutiae visualised on the extrac-ted vein pattern; (b) complex-modulus SML Fourier spectrum sampled on a polar-logarithmic grid; (c)complex-modulus SMC Fourier spectrum sampled on a polar-linear grid. (d) real-valued SML Fourierspectrum sampled on a polar-logarithmic grid; (e) real-valued SMC Fourier spectrum sampled on apolar-linear grid.

Refer to figure 5.2 for examples of sampling the minutiae input visualized in 5.2a with thedifferent approaches: The magnitude of the polar-logarithmic SML is shown in figure 5.2b whilethe magnitude of the polar-linear SMC is shown in figure 5.2c. Figures 5.2d and 5.2e show thereal-valued grids of the same SML and SMC. Real-valued SMR lose their translation invariancecompared to the absolute-valued SMR, but are more stable on low-quality data with variousspurious minutiae.

Further, it is possible to tune the SMR sampling to more stable, but less detailed con-figuration by altering λmin and λmax [XVK+08]. λmin correspondents to the lower frequencies

30

of a SMR and λmax to the higher frequencies. To receive a more stable SMR in a high noiseenvironment, the λmax should be lowered to values between 0.45 and 0.55. Altering λmin shouldbe avoided since all experiments that used a λmin 0.1 achieved a notoriously bad biometricperformance.

5.3 NormalisationSince the SMR yields spectra with different energies depending on the number of minutiae persample, each spectrum has to be normalised to reach zero mean and unit energy. To normalisethe SMR, the following equation is used.

X = M−Mσ(M)2 (5.9)

By subtracting the mean (M) of the magnitude2 (M) and then dividing it by the standarddeviation (σ(M)2) of it, the resulting spectrum X is normalised to zero mean and unit energy.

5.4 Feature ReductionSampling the spectra on a N = 256 and M = 128 grid yields a N×M = 32768 sized featurevector. This large-scale feature vector introduces two drawbacks:

Storage Considering N×M = 32768 double-precision float (64bit) values, each template wouldtake 2097152bit = 256kB RAM or data storage.

Comparison Complexity Processing a N×M = 32768 sized feature vector is a large com-putational task and limits comparison speeds, especially with large-scale databases inbiometric identification scenarios.

To address these issues, the same authors of the SMR approaches introduced two feature-reduction approaches in [XVKA09]. Both base on well-known algorithms and are explained inthe following subsections.

5.4.1 Column Principal Component Analysis

Based on the idea of the well-known Principal Component Analysis (PCA) originally presentedin [Pea01], the CPCA is introduced as one feature reduction approach. The authors presenttwo problems by applying the original PCA to the SMR. First mentioned is the small samplesize problem introduced in [RJ+91]: to receive a reliable PCA feature reduction, it is necessaryto train the PCA with a large training set. It is unlikely possible to acquire a large training setin real-world scenarios. The second mentioned problem is the rotation invariance of the PCA:

2Or real valued part.

31

(a)

(b)

(c) (d)

Figure 5.3: Illustration of the feature reduction approaches: (a) input minutiae visualised on theextracted vein pattern; (b) complex-modulus SML Fourier spectrum sampled on a polar-logarithmicgrid; (c) feature reduction with Column Principal Component Analsysis (CPCA); (d) feature reductionwith Line Discrete Fourier Transformation (LDFT) applied on top (c).

when applying the PCA to a SMR, the resulting feature vector loses its rotation invariance andevery probe has to be pre-rotated to a fixed position for all references and probes.To avoid both problems, the PCA is only applied column-wise, resulting in the CPCA. Naïvelyspoken, the CPCA scrutinizes the input SMR and applies the Singular Value Decomposition(SVD) column-wise.The first problem is thus addressed by the number of columns (M) per SMR. Let X be theM×N sized 2-D feature vector received by the SMR, every column is translated in a newfeature vector −→z = (z1, . . . , zM)T , thus every spectrum consists of N feature vectors X =(−→z 1, . . . ,−→z N). When L samples of X (X 1, . . . ,X L) are present in the training set, LN = L×N

trainings vectors Z = [−→z 1, . . . ,−→z LN] are received. In other words, the spectra get appended on

the X-axis to another and every column is analysed.The second problem of keeping the rotation operator (shifts on the horizontal axis) is avoidedby only reducing features in the means of rows. Thus the information of every column is stillin the same column after the CPCA transformation and rotations can still be compensated byhorizontal shifts.Since the features are concentrated in the upper rows after applying the CPCA, the lower rowscan be removed, resulting in a N×MCP CA sized feature vector. According to the authors, theachieved reduction is stated up to 90% for SMO and up to 80% for SML with the CPCAapproach while maintaining the biometric performance. Figure 5.3c shows the CPCA of theSML with proposed settings in figure 5.3b with a small L = 4. The SMR with applied CPCA-reduction will further be referenced as SMR reduced with the CPCA feature-reduction (SMR-CPCA) respective SML reduced with the CPCA feature-reduction (SML-CPCA) and SMC-

32

CPCA.

5.4.2 Line Discrete Fourier Transform

The second introduced feature reduction approach reduces the features in the horizontal direc-tion. It bases on the periodic characteristic of the horizontal axis of the spectrum X . Thus, itis possible to apply it on top of the CPCA reduction.In the LDFT feature reduction, every row of X is translated in a line vector −→y = (y1, . . . , yM),therefore X = (−→y 1, . . . ,−→y M)T . With the discrete Fourier transform [OWN96],X LDF T = (Y1(k), . . . , YM(k))T , an exact representation of X , is received. Only the Fourier com-ponents with a certain percentage of energy, the authors propose 80% (refer to [XVKA09] fordetails), are retained, and thus the feature reduction is achieved. The circular shift rotation isthen expressed as

T (m, n− ncs) = exp(−j2π

Nkncs

)×X LDF T ; k = 0, 1, . . . , N − 1; m = 1, . . . , M (5.10)

with ncs expressing the columns to shift.The authors state a reduction of up to 70% for the stand-alone LDFT and up to 95% for LDFTafter CPCA while maintaining the biometric performance.

5.5 BinarisationSince ∑Z

i=1 exp(−j(wxxi + wyyi)) ∈ C, thus M(wx, wy) ∈ R, every element in X is definedas a 32bit or 64bit floating-point real-valued number. Comparisons or calculations (especiallydivisions) with single or double precision floating-points is a relatively complex task comparedto integer operations. In the proposed SMR scenario, where a direct comparison (see section5.6) is defined as one multiplication, one division and one addition per feature, N×M×OPcnt =256×128×3 = 98304 floating-point operations3 are needed for one comparison. To address thiscomputational complexity and comply with other template-protection or indexing approacheswhere a binary feature-vector is required, the SMR can be converted to a binary feature vectoras presented in [XV10a].The authors introduce two binarisation approaches, each using a different data model but bothresulting in the same representation.

The binarisation approaches yield two binary vectors: a so-called sign-bit vector and a so-called mask-bit vector:

sign-bit The sign-bit vector contains the actual features of the SMR in a binary representation.Each bit is set according to one of the two binarisation approaches.

3A small additional overhead for preprocessing optimised values is ignored.

33

(a) (b) (c)

(d) (e) (f)

Figure 5.4: Input and result of the SMR-binarisation for SML and SML-CPCA: (a) real-valued SMLinput; (d) real-valued SML-CPCA input; (b) the spectral sign-bit of (a); (e) the spectral sign-bit of(d); (c) the spectral mask-bit of (a); (f) the spectral mask-bit of (d).

mask-bit Since binary representations suffer from bit-flips on edges in fuzzy environments,a second vector is introduced. This vector masks the likely-to-be-stable, called reliable,sign-bits.

The mask contained in the mask-bit vector is not applied to the sign-bit. Instead, it is kept asauxilary data and is applied during the comparison step4.The authors describe two approaches of receiving both vectors:

Spectral-Bits The Spectral-Bits are the result of the first, naïve approach. Let X R be theSMR feature vector, X BSsign the sign-bit vector and X BSmask the mask-bit vector of theSpectral-Bits. Then BX sign

is defined as

X BSsign(x, y) =

1, if X R(x, y) > 0

0, otherwise(5.11)

and Bmask as

X BSmask(x, y) =

1, if |X R(x, y)| > MT

0, otherwise(5.12)

Using the proposed threshold of 0.8 as MT for the mask-bit ensures masking only thestable bits of X BSsign.

Phase-Bits While the Spectral-Bits only evaluate the real-valued SMR, the Phase-Bits alsouse the imaginary part of the SMR spectrum X C. A third dimension is added to the sign-bit and mask-bit vectors for the Phase-Bits. The subsequent steps equal the Spectral-Bits approach but with an additional dimension. Let X C be described as X C(x, y) =(real + i ∗ imag) and again X BP sign the sign-bit vector and X BP mask the mask-bit vector

4This approach equals the masking procedure in iris recognition (see [Dau03, Dau04]).

34

of the Phase-Bits, then

X BP sign(x, y) =

[0, 0], if X C(x, y)real ≤ 0 and X C(x, y)imag ≤ 0

[0, 1], if X C(x, y)real ≤ 0 and X C(x, y)imag > 0

[1, 0], if X C(x, y)real > 0 and X C(x, y)imag ≤ 0

[1, 1], if X C(x, y)real > 0 and X C(x, y)imag > 0

(5.13)

and

X BP mask(x, y) =

[0, 0], if X C(x, y)real ≤ MT and X C(x, y)imag ≤ MT

[0, 1], if X C(x, y)real ≤ MT and X C(x, y)imag > MT

[1, 0], if X C(x, y)real > MT and X C(x, y)imag ≤ MT

[1, 1], if X C(x, y)real > MT and X C(x, y)imag > MT

(5.14)

For the mask-bit vector of the Phase-Bits, a threshold of 1.2 is proposed.

5.6 ComparisonIn the initial publication [XVK+08], two comparison methods for the SMR are introduced:one direct comparison approach and one weighted comparison approach. Both of them showedalmost identical results, and thus the weighted comparison approach was dropped in later pub-lications because the slightly better results did not justify additional training and computationaloverhead. Later on in [XVB+09], a third comparison method that tries to compensate rotationand scaling variances by using an additional Fourier transform is introduced. However, thisapproach shows a significantly lower performance in terms of FMR, FNMR and Equal ErrorRate (EER) and was also dropped in later publications. Therefore, the next sections will onlycover the direct comparison methods, that are used later in this thesis.

5.6.1 Spectral Minutiae

The most proven performance in SMR comparison is reached with the so-called direct com-parison5. It yields the most reliable comparison scores while keeping a minimal computationalcomplexity.Let R(m, n) be the spectrum of the reference template and P (m, n) the spectrum of the probetemplate, both sampled on the polar-logarithmic grid and normalised to zero mean and unitenergy (see section 5.3). Then, the similarity-score E

(R,P )DM is defined as

E(R,P )DM = 1

MN

∑R(m, n)P (m, n) (5.15)

The score is thus defined by correlation, which is a common approach in image processing.5In [XVK+08], the comparison method is named direct matching, where matching is used as wrong term for

comparison.

35

5.6.2 Binary Spectral Minutiae

For comparing two binary SMR or SMR-CPCA, a different approach is introduced in [XV10a],which is also used in the iris modality [Dau03, Dau04].After converting R(m, n) and P (m, n) in their individual mask-bit and sign-bit, yielding {maskR, signR}and {maskP, signP}, the fractional Hamming Distance (FHD) can be applied on those binaryrepresentations.

FHD(R,P ) = ∥(signR⊗ signP ) ∩maskR ∩maskP∥∥maskR ∩maskR∥

(5.16)

The inclusion of masks in the Hamming Distance masks out any likely-to-flip bits and onlycompares the parts of the sign-bit vector where the mask-bit vectors overlap. Therefore, onlythe likely-to-be-stable, so-called reliable, areas are compared. This typically improves the re-cognition performance.

5.7 Extending the Spectral Minutiae Representation withquality data

As an enhancement for the SMR, a approach of using quality information about minutiae ispresented in [XV09]. The authors differ between two types of quality:

Location Accuracy (qL) Under certain circumstances, the position of a minutiae cannot beaccurately determined. For example, when wide veins need to be thinned to their skeleton,a displacement of the actual vein centre could happen. The Location Accuracy describesthe expected certainty that the minutiae has not been misplaced in the feature-extractionprocess.

Minutiae Reliability (qM) It is possible for feature-extraction pipelines to falsely extractminutiae. Some pipelines are able to determine a genuine certainty for each minutiae thatdescribes the certainty that the extracted reference point is a genuine minutiae and nota spurious minutiae. The following steps can decide whether they use a minutiae or howto weight them.

Both types of quality can be incorporated independent of one another in all previously-presentedSMR, yieling the quality data enhanced Spectral Minutia Location Representation (QSML) andthe quality data enhanced Spectral Minutia Complex Representation (QSMC).Recall section 5.1.1, where equation 5.5 defines the Gaussian smoothed SML. The Gaussianfilter helps to reduce small location errors for all minutiae. Now, if the accuracy of minutiaelocations is known, instead of using a fixed σ Gaussian kernel for all minutiae, σ is lineardepending on the minutiae quality individually, thus yielding

M(wx, wy) =∣∣∣∣∣

Z∑i=1

(exp

(−

w2x + w2

y

2σ−2i

)∗ exp(−j(wxxi + wyyi))

)∣∣∣∣∣ (5.17)

36

as a new equation for the SML with location accuracy quality.When the minutiae reliability is known, the Dirac pulse (eq. 5.1) of each minutiae can beweighted linearly:

M(wx, wy; σ2) =∣∣∣∣∣exp

(−

w2x + w2

y

2σ−2

)Z∑

i=1wiexp(−j(wxxi + wyyi))

∣∣∣∣∣ (5.18)

A higher reliability correspondents to a higher weight wi for minutiae mi(x, y, qM).

-30

-20

-10

100250

0

Z -

Valu

e

10

200

Y - Rows

20

150

X - Columns

50

30

100

50

0 0

(a) Raw SMR

-30

-20

-10

100250

0

Z -

Valu

e

10

200

Y - Rows

20

150

X - Columns

50

30

100

50

0 0

(b) Location accuracy

-30

-20

-10

100250

0

Z -

Valu

e

10

200

Y - Rows

20

150

X - Columns

50

30

100

50

0 0

(c) Minutiae reliability

Figure 5.5: Effects of the quality data on the (not normalised) SML

Compare figure 5.5a and 5.5b: in 5.5b the location accuracy quality data is used with anaverage accuracy of 50%. The location accuracy smooths the high frequency parts of the SMRthat are sampled in the higher Y-axis regions (Y > 64), and thus the impact6 of a minutiawith a low accuracy is reduced in high frequencies, while minutia with a high location accuracykeep their impact on the high frequencies. With increasing frequency, a minutiae’s location is

6Increase of the SMR value (Z-axis).

37

narrowed down: the higher the frequency, the more precise the location information.By contrast, figure 5.5c shows the effect of the minutiae reliability on the SMR, compared tofigure 5.5a. The minutiae reliability reduces the impact on all frequencies of the SMR. Therefore,minutiae with a low reliability will leave a smaller footprint in the SMR than minutiae with ahigh reliability. Observe the lower frequencies along the beginning and the end of the X-axis,whereby the value of the SMR is reduced since unreliable minutiae have a lesser impact.Note: the QSML shown in figure 5.5 are not normalised to [−1, 1]. The normalisation processaccentuates the visualisation of the quality data impact since it evenly distributes the Z-axis.

5.7.1 Retrieving and applying minutiae-reliability data in vascularmodalities

Applying minutiae-reliability quality data promises increasing performance in terms of GAR. Inchapter 4, a feature-extraction pipeline was introduced which is already able to report minutiae-reliabilty data.

One approach for extracting quality data could be to use the contrast of the veins. However,this introduces at least one concern: image contrast is relative. To use the contrast as qualitydata, a reference value is needed. Using the average image brightness as reference could yieldsmall contrast gradations. Second, it is very hard to normalise the output of the contrast qualitydata approach. Therefore, another quality data extraction approach is proposed.

Recall section 4.4, whereby the maximum-curvature approach yields a score map describingthe probability that a local maxima actually represents the centre of a vein. Observe figure

(a) (b) (c)

Figure 5.6: Feature detection with the maximum-curvature approach: (a) input probe; (b) response ofthe maximum-curvature algorithm; (c) probe (a) with layover response (b).

5.6, it is visible, that the response also corresponds to the darkness and contrast of the vein:clearly-visible veins result in a high response by the maximum-curvature algorithm, while less

38

noticeable veins result in a low response. Since these small, less noticeable veins are more likelyto be affected by short-term influences like temperature or blood pressure, it is desirable toweight their minutiae less than those of the larger veins. As described in section 4.5, the usedthinning algorithm best keeps the initial shape and best hits the actual centre of the input.Thus, it is safe to assume that the minutiae found in the skeleton will be in the centre oftheir maximum-curvature extracted veins. Following the assumption, the value at the minutiae

(a) (b)

(c) (d)

Figure 5.7: Extending the minutiae with reliability data taken from the response of the maximum-curvature algorithm: (a) response of the maximum-curvature algorithm of the probe showed in figure5.6a (jet colour-map); (b) skeleton of the vein pattern; (c) scored minutiae (red = lowest, green =highest score); (d) fusion of the previous figures (maximum-curvature without jet colour-map; minutiaeyellow = lowest, green = highest score).

location in the maximum-curvature response corresponds to the minutiae reliability.A drawback of this approach is its dependency on overall contrast: probes with low contrast,even after the image enhancement, will automatically result in a overall lower reliability. Toaddress this issue, the maximum-curvature response has to be normalised to [0, 1]. Let −−→MC

be the maximum-curvature response as one continuous vector, then the normalised maximum-

39

curvature vector MC is given by

MC(x, y) = min(

gMC(x, y)max

−−→MC

, 1)

(5.19)

where g is a gain factor used to adjust the minimal needed response for 100% (MC(x, y) = 1)minutiae reliability. After applying the normalisation to the maximum-curvature response, thevalues are contrast-invariant and can safely be used as reliability descriptors.For further confidence, it is possible to include the minutiae N4- or N8-neighbourhood of thenormalised maximum-curvature response in the calculation and take the mean or weighted meanas minutiae reliability. Thus scores of minutiae detected from single pixel width maximum-curvature responses will be reduced while the score of reliable thick responses will barely bealtered.

5.7.2 Retrieving and applying location-accuracy data in vascularmodalities

Even if the thinning algorithm used (see section 4.5) promises the best skeleton accuracy, it ispossible for the skeleton to be shifted by a few pixels. Therefore, it might benefit the GAR toretrieve and use the location-accuracy data to reduce errors introduced by shifted minutiae.Again using the maximum-curvature response, the location-accuracy can be approximated forindividual minutiae points. Instead of using the normalised maximum-curvature response, abinary representation will be used. The location accuracy is then approximated by the widthof the vein. A naïve approach is to test the vein width by taking the minimum span of setbits of the binary maximum-curvature response MCB in each direction as proposed in listing1. The returned pixel span can now be used to approximate the location accuracy relativelyto the probes size: minutiae where their returned SpanDir0 differs from SpanDir1 by morethan a few pixel are likely to be shifted in other probes. Thus, the difference of SpanDir0 andSpanDir1 (∆SpanDir) corresponds linearly7 to the location accuracy.

5.8 Minutiae pre-selectionBesides the quality extensions presented in the previous section, the SMR can further be expan-ded by a simple minutiae pre-selection. Since the feature extraction pipeline extracts variousspurious minutiae, it holds interest whether the biometric performance can be increased byautomatically ignoring minutiae, based on simple rules:

Type A naïve selection approach is to either select only endpoints or select only bifurcationsfor the SMR generation. However, it is expected that only selecting endpoints results

7Note that the difference has to be put in perspective to the probe’s size: a difference of 5px is much less ofan issue in a 512×512 sized probe than in a 128×128 sized probe.

40

Algorithm 1 Naïve vein span algorithm for minutiae m(x, y).1: function VeinSpan(MCB, x, y)2: sHeight0 ← 1, sHeight1 ← 1, sWidth0 ← 1, sWidth1 ← 1, sDiagA0 ← 1, sDiagA1 ←

1, sDiagB0 ← 1, sDiagB1 ← 13: grow ← true, growFactor ← 14: while grow = true do5: heightGrow ← 0 ▷ Height-Span6: if MCB(x, y + growFactor) = 1 then7: sHeight0 ← sHeight0 + 1, heightGrow ← heightGrow + 18: end if9: if MCB(x, y − growFactor) = 1 then10: sHeight1 ← sHeight1 + 1, heightGrow ← heightGrow + 111: end if12: widthGrow ← 0 ▷ Width-Span13: if MCB(x + growFactor, y) = 1 then14: sWidth0 ← sWidth0 + 1, widthGrow ← widthGrow + 115: end if16: if MCB(x− growFactor, y) = 1 then17: sWidth1 ← sWidth1 + 1, widthGrow ← widthGrow + 118: end if19: diagAGrow ← 0 ▷ Diagonal-Span (/)20: if MCB(x + growFactor, y − growFactor) = 1 then21: sDiagA0 ← sDiagA0 + 1, diagAGrow ← diagAGrow + 122: end if23: if MCB(x− growFactor, y + growFactor) = 1 then24: sDiagA1 ← sDiagA1 + 1, diagAGrow ← diagAGrow + 125: end if26: diagBGrow ← 0 ▷ Diagonal-Span (\)27: if MCB(x− growFactor, y − growFactor) = 1 then28: sDiagB0 ← sDiagB0 + 1, diagBGrow ← diagBGrow + 129: end if30: if MCB(x + growFactor, y + growFactor) = 1 then31: sDiagB1 ← sDiagB1 + 1, diagBGrow ← diagBGrow + 132: end if33: if heightGrow = 0 or widthGrow = 0 or diagAGrow = 0 or diagBGrow = 0

then34: grow ← false

35: end if36: growFactor ← growFactor + 137: end while38: return min ((sHeight0, sHeight1), (sWidth0, sWidth1), (sDiagA0, sDiagA1), (sDiagB0, sDiagB1))39: end function

41

in very poor biometric performances since only selecting endpoints corresponds to onlyselecting noise.

Reliability Based on the minutiae-reliability extraction presented in section 5.7.1, only minu-tiae with a reliability score that exceeds a certain threshold will be selected for SMRgeneration.

Neighbourhood The neighbourhood selection (neighbourhood cleaning) extends the approachto only select bifurcations: all endpoints are excluded from the minutiae list. Further, thebifurcation minutiae whose closest neighbour (in the means of Euclidean distance) is aendpoint minutiae are also excluded.

The reliability criterium can be combined with the other two, resulting in five minutiae selectionapproaches. Applying one or a combination of the minutiae pre-selection to the SMR yieldsthe Spectral Minutia Representation with minutiae pre-selection (PSMR), respective the PSMRsub-types Spectral Minutia Location Representation with minutiae pre-selection (PSML), Spec-tral Minutia Complex Representation with minutiae pre-selection (PSMC), PQSML and PQSMC.

5.9 Chapter ConclusionsIn the first section of the chapter, the SMR has been outlined based on the original publicationsand the collection in [XV10c]. With the SMR, Xu et al. introduced a promising representationfor templates of fingerprints. The SMR features comparable characteristics for fingerprint tem-plates like the Iris-Code (see [Dau04]) for iris recognition: yielding a fixed-length feature vectorwith template-protection properties and fast comparison methods for the complex-valued aswell as the binary-valued representation.Since the SMR is based upon minutiae data, it should be feasible to apply the SMR to otherminutiae-based modalities. Hartung et al. analysed the feasibility of the SMR for finger veins,wrist veins and veins in the back of the hand in [HOXB11] and [HOX+12]. The authors con-cluded the feasibility of the SML variant for those vein modalities. Continuing on the work ofHartung et al., the experiments in section 9.2 analyse the feasibility of the SMR for the palmvein characteristic and state recommendations and hints about applying the SMR in the palmvein modality.The final two sections outlined the quality enhancements for the SMR by [XV09] with the SMLas a example with a new contribution: a proposal to retrieve and apply minutiae-reliabilityas well as minutiae-accuracy data based on the maximum-curvature approach. Enhancing theSML with quality data yieldes the QSML, and enhancing the SMC yields the QSMC repres-entation. The approaches are expected to raise the GAR for vascular minutiae and are subjectto experiments in section 9.2.Thus far, no approaches for workload reduction to address the issue stated in section 2.2.3 have

42

been introduced. The next chapters introduce two promising indexing approachs for workloadreduction that are used in the experiments.

43

Chapter 6

Bloom filter-based indexing

In section 2.2.3, the challenges associated with large-scale biometric identification databases aredescribed. To tackle those challenges, workload reduction strategies have to be implemented.This chapter outlines the first workload reduction approach chosen in this thesis using theBloom filter.

Data Capture Subsystem

Signal Processing Subsystem

Data Storage Subsystem

Comparison Subsystem

Decision Subsystem

Template

Database

Template selection

Comparison

Candidate list

Figure 6.1: Bloom filter-based indexing

6.1 Bloom FilterIntroduced as early as 1970 in [Blo70], the Bloom filter is still used today in a broad scopein computer sciences (e.g. Big Data [CDG+08], CDN-caching [MS15], URL-matching [Yak10,Hes12] and computer forensics). The purpose of the probabilistic data structure is an efficientmembership query. A Bloom filter (denoted B) is organized as a vector of l bits. In the initial(empty) Bloom filter, every bit is set to zero.Upon adding or for querying, an element has to be converted to its Bloom filter representation.

44

Hash-set creation The new element or query is hashed by a fixed number (k) of independenthashing functions H1, . . . , Hk that produce hash values in the range of [0, l] resulting inthe hashes h1, . . . , hk. These hash values represent indices for the Bloom filter.

Now the hashes h1, . . . , hk can either be added to the Bloom filter or a membership query canbe run:

Adding to the Bloom filter An element is added to the Bloom filter by taking the hashvalues hi as indices for the Bloom filter and setting the corresponding bits to one as byBN [hi] = 1 ∈ 0 . . . k. Thus an element is added to B by B = B ∨ BN .

Membership query As well as in adding to the Bloom filter, the queried element is convertedto its Bloom filter representation by BQ[hi] = 1 ∈ 0 . . . k. Now the query is performed bycomparing the Bloom filter with the query by |B∧BQ|. Only if |B∧BQ| = |BQ| is fulfilledis the query’s element contained in the Bloom filter. Thus, if any of the set bits in BQ isnot set in B, the element is definitely not part of the Bloom filter, and therefore no falsenegatives occur. However, it is possible for false positives to occur due to bit collision.Therefore, a Bloom filter is only able to tell if the element is perhaps or definitely notpart of the Bloom filter.

Bit collisions cannot be prevented, but can be delayed by increasing l. A Bloom filter is full ifall bits are set and thus all queries will yield a positive result. It is desirable to keep a Bloomfilter as a sparse matrix to minimise the probability of bit collisions. The event of exceedingthe threshold of set bits for a Bloom filter is further called overflow in this thesis.

In its basic definition, the Bloom filter offers very successful and efficient membership queriesfor scenarios where the probe and the reference are identical. However, in biometric systems,the Bloom filters can be enabled to work with fuzzy data, allowing its usage in verification andnaïve identification scenarios if an appropriate comparison function (e.g. the hamming distance)is used.The initial objective of this work was to apply the Bloom filter on the SMR to achieve asignificant workload reduction for identification scenarios. In [RBB13] and [RBBB15], the Bloomfilter approach is already successfully applied to Iris-Codes. Since Iris-Codes are fixed-length,binary structures, the approach is tailored to templates representing thsee characteristics. Aspresented in section 5.5, the SMR is also transferable in a fixed-length binary representation ofbiometric references. Therefore, applying the same concept to the binary SMR seems to be areasonable approach. In the next sections, the basics of this concept are outlined.

6.2 System BasicsIn this section, operational details of the Bloom filter-based scheme presented in [RBBB15] andextended by [DRB17] will be presented. Both publications provide the basis and the baseline for

45

the workload reduction approach, which is benchmarked against a new contribution presentedlater in this thesis.

6.2.1 Template Transformation

To transform a SMR template to its Bloom filter representation, the procedure explained belowis used.

1. The SMR has to be converted to its binary form (see section 5.5). Since the basic formdoes not support auxiliary-data schemes, i.e. the sign, mask pair, the mask-bit has to beapplied to the sign-bit so that matrix = sign ∧mask.

2. The resulting binary matrix has to be segmented into equally-sized blocks of width Wand height H.

3. In contrast to the original Bloom filter approach, where multiple hash function are appliedto the data, a simple mapping scheme is used. Each block is mapped to a Bloom filter(B) by interpreting the bits of each column (c1, . . . , cW) as unsigned integer (h1, . . . , hH),which are then used as indices for setting the bits in the Bloom filter. This representsan inverted hashing strategy: the original Bloom filter uses multiple hashing functions ona single element, whereas here multiple elements (the columns) are hashed by a singlehashing function. Both strategies yield the same end result of h1, . . . , hW hash valuessubsequently used as indices.

4. Repeat step 3 for every block: each block yields a Bloom filter Bi resulting in a sequenceof Bloom filters (B = {B1, . . . ,BBC}) of length BC = N

W×MH .

The transformation features rotation-compensating properties for the binary SMR matrix. Inmost width and height configurations, fewer shifts for rotation compensation have to be done,due to the indices scheme.The size (τ) of a Bloom filter sequence template with block sizeW×H for a N×M binary SMRmatrix is calculated by equation 6.1.

τ = 2H ∗ N

W∗ M

H(6.1)

With the described template transformation, it is possible to use the Bloom filter in a verificationand a naïve identification scenario. A sequence of Bloom filter for a N = 256×M = 128 binarySMR with a block size of W = 8×H = 4 reduces the bit comparisons by 50%.

However, in identification scenarios the overhead for large-scale databases is dwindling re-lative to the overall computation time. To use the modified Bloom filter for a more efficientidentification than a naïve approach, the methods described in the following two sections areused.

46

6.2.2 Tree Construction

Storing the Bloom filter-based templates in a binary search tree enables a more efficient iden-tification by reducing the number of comparisons needed. The process of building the tree isdescribed below and conceptually illustrated in figure 6.2.

‚

‚

‚

‚

‚

‚

B

B

B

B

B‚

B

B

B

‚

S

Ui=1

Bi

S/2

Ui=1

Bi

S

Ui=S/2+1

Bi

S/4

Ui=1

Bi

S/2

Ui=S/4+1

Bi

SI(leƒt, )

SI(right, )>

rightleft

rightleft

. . .

. . .

SI(leƒ t, )

SI(right, )>

. . .

SI(B1, )

SI(B2, )≤B1 B2 . . .

. . .

BS-1 BS

retrievalindexinglist of NbinarySMR

SI(B2, )?> threshold return or B2 Ø

Figure 6.2: Bloom filter tree construction scheme and tree traversal for query B′ (taken from [DRB17],original in [RBBB15]).

1. The tree root is created as the element-wise union of all templates (S) such that Ψ0 =B1 ∪B2 ∪ · · · ∪BS =

S∪i=1

Bi.

2. Child nodes are created by splitting the template list of their parents in two parts andtaking the element-wise union of one of the two sub-sets (i.e. Ψ1 = B1 ∪ · · · ∪ BS/2 =S/2∪i=1

, Ψ2 = B(S/2)+1 ∪ · · · ∪BS =S∪

i=S/2+1).

3. Repeating step 2 until only the leafs are left, which then are the templates (B1, . . . , BS)themselves (e.g. Ψ2S−2 = BS , if S is a power of two).

Inserting a new template into an already-built tree is trivial and can be achieved in O(logS)steps: The new template is inserted as a new leaf and is added to the element-wise union of itsparents and their predecessors. If no node has an empty leaf, a leaf is converted to a node and

47

receives its template as well as the new template as children (respectively leafs). Removing oneor more templates requires fully rebuilding1 the tree, which is done in O(SlogS) steps.

6.2.3 Tree Traversal

Upon receiving a membership query, the query is transformed in its Bloom filter representation(B′). Subsequently, the tree traversal for a single tree starts at the root node (Ψcurrent =Ψ0) by computing the similarity score for the two children (Ψcurrent+1, Ψcurrent+2) of the rootnode. Basing on the two received scores, the direction of the first descend is determined, thusΨcurrent = Ψcurrent+1 if the similarity score of Ψcurrent+1 and B′ is higher than the score ofΨcurrent+2 and B′, respectively Ψcurrent = Ψcurrent+2 vice versa. These direction decisions arerepeated until a leaf is reached. Finally, the similarity score of the leaf and B′ is compared to athreshold, which has been previously trained with a set disjoint from the enrolled templates ormanually set by an operator. The lookup process is conceptually illustrated in figure 6.2 (boldlines) and formally stated in algorithm 2.The similarity score of two Bloom filter sequences (B, B′) is calculated by a score-level fusionof the pair-wise similarities between the Bi ∈ B and B′

i ∈ B′ as shown in function 6.2. Bytaking the amount of matching bits and the population count of the set of matching bits ofbooth B and B′, the similarity is calculated. Because the population counts of B and B′ varies,it is necessary to normalise the score by subtracting the average of both population counts.Equation 6.3 shows the final similarity calculation for B and B′.

SI(B, B′) = 1BC

BC∑i=1

SI(Bi,B′i) (6.2)

SI(B,B′) = |B ∧ B′|12(|B|+ |B′|)

(6.3)

Because a decreasing similarity score per level is expected for impostor queries, an additionalworkload reduction, for impostor queries and queries for templates stored in another tree, canbe achieved by prematurely stopping the tree traversal for the current tree, when the similarityscore of the different levels is required to be ordered (i.e. SIlevel0 < SIlevel1 < SIlevel2 < . . . ).As a side effect, it is more difficult for impostors to be accepted, thus potentially lowering theFalse Positive Identification Rate (FPIR). However, some genuine queries may also be rejected,potentially increasing the False Negative Identification Rate (FNIR). Therefore, requiring aincreasing similarity per level is considered as optional and is only recommended for biometricsubsystems with a very robust feature-extraction subsystem.

By increasing the number of stored templates (S) in a Bloom filter tree, the populationcount2 of the root nodes B(Ψ0) increases. While this is less an issue for the root node itself, an

1A derivative of the Bloom filter, the so-called counting Bloom filter (see [RB13]) addresses this issue on thecost of the efficiency of performing bitwise operations.

2In other words, the sparsity of the root node decreases.

48

excessive population count in the lower nodes affects the capability of making correct traversaldirection decisions. A wrong traversal direction decision at the upper levels resulting fromtoo many bit collisions is not correctable in lower levels, thus resulting in a False NegativeIdentification (FNI) event or worse in a False Positve Identification (FPI) event3. Therefore,it is necessary to build multiple trees with fewer subjects respectively biometric instances pertree. The lookup process for multiple trees employs the same strategy for traversing a tree asfor single tree environments. Merely a trivial loop-logic has to be added, as shown formally inalgorithm 3.

Algorithm 2 Tree traversal for a single Bloom filter tree.1: function TraverseTree(Node, Query, Threshold, LastScore)2: if Node[left] = nil and Node[right] = nil then3: scoreLeft← similarity(Node[left], Query)4: scoreRight← similarity(Node[right], Query)5: if scoreLeft > scoreRight then6: if scoreLeft < LastScore then7: return nil ▷ Optional8: end if9: return TraverseTree(Node[left], Query, Threshold, scoreLeft)10: else11: if scoreRight < LastScore then12: return nil ▷ Optional13: end if14: return TraverseTree(Node[right], Query, Threshold, scoreRight)15: end if16: end if17: score← similarity(Node[this], Query)18: if score > Threshold and score > LastScore then ▷ Second condition optional19: return Node[this]20: end if21: return nil

22: end function

6.2.4 Configuration

By now, three variables are to be considered for the basic Bloom filter tree system: Bloomfilter width (W), Bloom filter height (H) and tree count (T ). Each of them can be adjustedconsidering the following limitations.

W Increasing the block width for the Bloom filter increases the number of bit collisions whentransforming the binary SMR-matrix to the Bloom filter representation. Reducing the

3This behaviour was observed in the experiments: queries that would normally have the highest similarityscore with their corresponding reference template in a naïve approach do not find their correct reference in theBloom filter since the high amount of bit collisions direct the query to a different reference template where thesimilarity score unfortunately exceeds the decision threshold.

49

Algorithm 3 Additional loop-logic for multiple Bloom filter trees.1: function Lookup(Trees, Query, Threshold)2: candidates← empty list3: for all Trees as root do4: candidate← TraverseTree(root, Query, Threshold)5: if candidate = nil then6: candidates← candidates + candidate

7: end if8: end for9: return candidates

10: end function

block width increases the total template size, thus requiring more RAM, storage and bitcomparisons. Further, it diminishes the rotation invariance properties of the templates.

H Increasing the block height for the Bloom filter increases the total template size. Reducingthe block height increases the number of bit collisions and increases the overall populationcount of the Bloom filter.

T Employing additional trees for the Bloom filter reduces the number of templates per tree,therefore reducing the bit collisions and increasing the hit rate, thus decreasing the FNIR.However, additional trees require more template comparisons, thus increasing the work-load. Refer to figure 6.3, with S = 512 using two trees nearly doubles the number oftemplate comparisons needed where using 64 trees hardly yields a workload reduction.

6.3 State-Of-The-Art Bloom filter approachAs already mentioned in section 6.2.2, wrong traversal decisions can occur if the child nodes aretoo populated4, thus it is required to build multiple trees with fewer subjects. However, doingso introduces a new issue: multiple trees cause higher workload. For a single tree environment,the number of template comparisons needed per tree is given by 2log2(S)− 1 (two comparisonsper level). Splitting the same number of subjects to multiple trees T , the number of templatecomparisons needed is

C = T ∗(

2 ∗ log2

(ST

)). (6.4)

Note figure 6.3, for an example with S = 512, 18 comparisons are needed in a single-treeenvironment, while using two trees nearly doubles the number of comparisons needed andusing 64 trees nearly invalidates the Bloom filter tree approach. In the basic setup, every treeis traversed and there is no approach to skip trees that unlikely contain the searched template.

4It isn’t possible to generally state a threshold when a child node is too populated since it depends on theconfiguration of the Bloom filter size and the number of bit errors in the raw binary templates.

50

1 2 4 8 16 32 64 12816

32

64

128

256

512

Trees built (T )

Com

paris

ons

(C)

T (2log2(512T ))

Figure 6.3: Number of needed comparisons (C) per lookup for S = 512 with 1, . . . , T trees.

6.3.1 Tree selection

Recall that the additional workload introduced by employing multiple trees reduces the workload-reducing effect of the Bloom filter-indexing. The estimated number of template comparisons(C) for single- and multi-tree environments is summarized in equation 6.5.

C =

2 ∗ log2(S)− 1 if T = 1

T ∗(

2 ∗(

log2

(ST

)− 1

))if 1 < T <

S2

S if T ≥ S2

(6.5)

A naïve approach to reduce the impact of additional trees would be to stop traversing thetrees, in a one-to-first search manner, if an appropriate candidate has been found. However,this naïve approach only reduces the number of template comparisons needed by factor 2 onaverage5

C = T2

(2 ∗

(log2

(ST

)− 1

)). (6.6)

A more efficient strategy is presented in [DRB17]. The idea is to add a pre-selection step,selecting the t most promising trees with t ≪ T . To determine the most promising trees, thequery template is compared with the root node of each tree. Only the t trees with the highestsimilarity between the root node and query template are then fully traversed. This concept isformally described in listing 4.

Note equation 6.7, using this strategy greatly reduces the additional workload introducedby the additional trees.

C =

2 ∗ (log2(S)− 1) if T = 1

T + t ∗(

2 ∗(

log2

(ST

)− 1

))if 1 < T <

S2

(6.7)

However, as shown in figure 6.4, the pre-selection scheme loses its efficiency when the numberof trees T approaches the number of templates S: The generated overhead by comparing the

5I has to be noted, that this is only true for genuine transactions, for impostor attempts, the workload isfurther reduced.

51

Algorithm 4 Improved loop-logic for multiple Bloom filter trees.1: function Lookup(Trees, t, Query, Threshold)2: rootScored← empty list3: for all Trees as root do4: rootScored← similarity(root, Query) + rootScored5: end for6: SortAscending(rootScores)7: candidates← empty list8: for i = 0 . . . t do9: candidates← TraverseTree(rootScored[i], Query, Threshold)10: end for11: return candidates12: end function

query template to the root nodes starts to overweight, thus reducing the overall workloadreduction.

100 101 102 103 104 105101

102

103

104

105

T

C(S

=218

)

BaselineBasic schemeStop-after-candidate schemeImproved scheme t = 1Improved scheme t = 2Improved scheme t = 16Improved scheme t = 128

Figure 6.4: Number of needed template comparisons (C) per lookup with the pre-selection scheme.

6.3.2 Quick traversal direction decision

Recall (section 6.2.3), in the basic system each traversal direction decision requires two compar-isons, thus yielding C = 2∗log2(S)−1. With the usage of the assumption stated in section 6.2.3,the similarity scores increases per level (for queries with their associated template included ina tree) when traversing a tree, the number of template comparisons needed can be reduced toone per level, thus C = log2(S)−1. Instead of comparing both children per node, only one (e.g.the left/first) children is compared. When the similarity score has not increased compared tothe previous, the other children is selected as next node; if the score has increased, the com-

52

pared node is selected. Thus, a second6 comparison is only needed if the similarity score has notincreased. Algorithm 5 formally shows the improved direction decision strategy. On average,

Algorithm 5 Tree traversal for a single Bloom filter tree with single comparison strategy.1: function TraverseTree(Node, Query, Threshold, LastScore)2: if Node[left] = nil and Node[right] = nil then3: scoreLeft← similarity(Node[left], Query)4: if scoreLeft > LastScore then5: return TraverseTree(Node[left], Query, Threshold, scoreLeft)6: else7: scoreRight← similarity(Node[right], Query)8: return TraverseTree(Node[right], Query, Threshold, scoreRight)9: end if10: end if11: score← similarity(Node[this], Query)12: if score > Threshold and score > LastScore then ▷ Second condition optional13: return Node[this]14: end if15: return nil16: end function

the second comparison is needed for every second level, yielding C = 32 log2(S)− 1.

Finally, it has to be noted that this improvement is not feasible if the assumption of increasingsimilarity scores (for queries with their associated template included in a tree) is not satisfied.

6.4 Tree root emulation and random template emulationIt is crucial to find a suiting tree-to-template ratio to receive an efficient workload reduction:using an insufficient number of trees increases the FNIR7 and is difficult to detect at runtime.For this purpose, it is necessary to emulate the root node of a tree and random templates.In [DRB17], Drozdowski presented an approach to mathematically emulate random templates,thus emulating root nodes of trees. By calculating the overlap between a emulated randomtemplate and the emulated root node, the author is able to estimate the number of trees tobuild for a database with S templates. With a few adaptations, this approach should also befeasible for binary SMR Bloom templates, too.Recall (section 6.2.1) that a Bloom filter B is created of a binary SMR block with width Wand height H. By assuming that all set bits in the block are mutually independent and areuniformly distributed, then:

B = {x ∈ N0 | 0 ≤ x ≤ 2H}, |B| = 1−(

1− 12H

)W(6.8)

6Refer to appended section A.1 for a excursus why a second comparison is needed.7It is possible to reach FNIR values that render the biometric system useless: With one or more levels

overflown, finding the right template is more a fluke than a hit.

53

Figure 6.5: Spectral mask-bit (figure 5.4c) applied to spectral sign-bit (figure 5.4b) of the SML infigure 5.4a as introduced in section 6.2.1.

Observe figure 6.5, where the spectral mask-bit has been applied to the sign-bit for theBloom filter transformation. It is clearly visible that for the SMR there is a very high dependencybetween individual columns and the bits set aren’t distributed uniformly. To allow for thisdeviation, let ε denote the difference between the expected number of duplicate columns in abinary SMR and randomly-generated, mutually-independent columns:

B = {x ∈ N0 | 0 ≤ x ≤ 2H}, |B| = 1−(

1− 12H

)W− ε (6.9)

There is no general value for ε since its value depends on parameters such as block size, thedataset itself, SMR configuration, SMR binarisation configuration and has small variationsbetween different templates.The concept can be extended for root nodes of the Bloom filter tree, too. Recall (section 6.2.2)that a Bloom filter tree root node (here RK) is a union of its (here K) child nodes, thus:

RK = B1 ∪ B2 ∪ · · · ∪ BK, |RK| =(

1−(

1− 12H

)K∗(W−ε))∗ 2H (6.10)

At this point, it is possible to approximately emulate random templates, thus emulatingrandom impostor templates and root nodes of Bloom filter trees. By estimating the overlap O ofthe tree root node (respective nodes) and an random impostor Bloom filter template (equation6.11), it is possible to qualify a statement about the feasibility of the current configuration.

O = |RK ∩ B| (6.11)

The expected overlap outcome follows a hypergeometric distribution with denoting the numberof overlapping items in range 1 . . . |B|:

P (O = o) =

(|B|o

)(2H−|B||RK |−o

)(

2H

|RK|

) (6.12)

The mean of that distribution Θ (equation 5) can be used as a single number metric.

Θ = |RK| ∗|B|2H (6.13)

Refer to [DRB17] for a explanation how to interpret the introduced Θ.

54

6.5 SummaryIn the first two sections of this chapter, the fundamental details of a Bloom filter-based biometricsystem have been outlined, partially based on the original article in which this approach wasproposed [RBBB15]. The third section outlines the state-of-the-art for the Bloom filter-indexingapproach by Drozdowski [DRB17]. All sections include minor adoptions to enable the approachfor the binary SMR.How well the adoptions and the Bloom filter itself perform in identification scenarios is subjectsto the experiments introduced in section 8.4 and evaluated in section 9.4.

55

Chapter 7

Spectral Minutiae CPCA-Tree indexing

In this chapter, a second workload reduction approach based on the SMR-CPCA is introduced.The system basics are adopted from chapter 6.

Data Capture Subsystem

Signal Processing Subsystem

Data Storage Subsystem

Comparison Subsystem

Decision Subsystem

Template

Database

Template selection

Comparison

Candidate list

Figure 7.1: Spectral Minutiae CPCA-Tree indexing

7.1 System BasicsIn contrast to the Bloom filter-based indexing approach, this author proposes a tree indexingscheme based on the SMR-specific CPCA. The CPCA already proved its feature-reduction, thusalso workload-reduction, capabilities without impairing the recognition performance1 [XVKA09].The basics of the SMR-CPCA are already introduced in section 5.4.1. However, for this chapter,a closer observation of the SMR-CPCA representation is reasonable.

1This statement by Xu et al. will also be tested for palm vein (respective vein based) SMR.

56

7.1.1 Binary SMR-CPCA merging

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

Figure 7.2: SML-CPCA creation from different subjects: (a,d,g) input SML; (b,e,h) CPCA represent-ation; (c,f,i) binary CPCA (top to bottom: sign-bit, mask-bit, applied mask-bit to sign-bit)

Observe figure 7.2, where three different real-valued SMR (here SML) and their (binary)SMR-CPCA are illustrated: the binary SMR-CPCA features a high inter-class variance - all setbits in the binary SMR-CPCA matrices are differently distributed and there are few unanimousbits.While a high inter-class variance is mandatory and desirable for a good recognition perform-

ance, it introduces some difficulties for naïve indexing schemes that rely on trivial stacking ormerging. When merging SMR-CPCA, the population count rises quickly, thus diminishing thedescriptive value. In other words, the bits set in a binary SMR-CPCA have few bit collisionswith SMR-CPCA from other subjects repectively from other biometric instances. The problemcan easily be explained visually: in figure 7.3, the bit-matrices for 1, 2, 4, 8 and 16 (randomlychosen) merged SMR-CPCA are shown and the percent of bits set are presented in table 7.1.Observe the quickly rising population count for the sign-bit (figure 7.3 left column) and mask-bit (figure 7.3 centre column). Already at 4 merged SMR-CPCA, the sign-bit and mask-bithave passed a population count where no descriptive value is left. If the mask-bit is appliedto the sign-bit and the result (applied-bit) is merged (figure 7.3 right column), then about 8SMR-CPCA can be merged without losing too much information, thus keeping the descriptivevalue. These conclusions will be further analysed in the experiments described in section 8.4.

57

(a) 1 (b) 1 (c) 1

(d) 2 (e) 2 (f) 2

(g) 4 (h) 4 (i) 4

(j) 8 (k) 8 (l) 8

(m) 16 (n) 16 (o) 16

Figure 7.3: SMR-CPCA (SML-CPCA) sign-bit (left), mask-bit (center) and applied mask-bit to sign-bit (right) for merged SMR-CPCA with 1, 2, 4, 8 and 16 SMR-CPCA (white = set, λmin = 0.1, λmax =0.6, MCP CA = 24).

No. merged Bits setSMR-CPCA sign-bit mask-bit applied-bit

1 50.3% 43.5% 21.4%2 74.5% 67.9% 38.1%4 93.3% 89.3% 61.4%8 99.3% 98.8% 83.9%

16 99.9% 99.9% 96.6%

Table 7.1: Average rate of set bits in the SMR-CPCA (SML-CPCA) sign-bit, mask-bit and appliedmask-bit to sign-bit for merged SMR-CPCA.

7.1.2 Masking out most common bits

Masking out bits that are set in a bulk of all SMR-CPCA templates (here SML-CPCA) isa common-known strategy used to lower the average similarity score of impostor comparisontrials: the impostor comparison trials that gain their score from these commonly set bits aremore likely to be rejected. However, since the CPCA transformation depends on the trainingdataset and the data itself, there is no common mask. As a rule of thumb, the upper bits are

Figure 7.4: Example heat map of set bits generated with 467 SMR-CPCA (SML-CPCA) applied-bit(red = most, blue = least).

more likely to occur in SMR-CPCA than the lower bits. Figure 7.4 shows an example heat map

58

of 467 (trainig-set size 33) SMR-CPCA generated with the PolyU-Database ([pol], see section8.1). Observe the upper rows of the heat map: these bits are set most often with a hot-spot inthe top-centre and top-right. Most descriptive value is held in the middle rows: these bits areset less often and more likely to differ between multiple SMR-CPCA. In this example scenario,it is recommended to mask out the upper three rows and the red hot-spots in the top-centreand right. The experiments in 8.4 will further analyse the advantages and disadvantages of thisstrategy.

7.1.3 Template Transformation

For the proposed SMR-CPCA binary search tree (CPCA-Tree) approach, no additional tem-plate transformation is needed. The SMR is merely transferred in its binary CPCA represent-ation and cut to MCP CA rows. Refer to section 5.4.1 for details.

7.1.4 Tree Construction

The tree construction for the proposed CPCA-Tree approach follows the same scheme as theBloom filter tree of section 6.2.2 with the Bloom filter templates replaced by SMR-CPCAtemplates. There are two different methods of representing the binary SMR-CPCA for theCPCA-Tree:

SMR-CPCA-Components (SMR-CPCA-C) Recall (section 5.5 and 5.4.1), there are twodifferent strategies when transforming a SMR in its binary form. The spectral binarisa-tion method yields two binary matrices: a sign-bit and a mask-bit. These two matricesrepresent the two components of the SMR-CPCA-Components (SMR-CPCA-C) repres-entation.

SMR-CPCA-Applied (SMR-CPCA-A) Instead of keeping both binary components, themask-bit is applied to the sign-bit, yielding the applied-bit (see section 7.1.1). The applied-bit matrix represents the single component representation SMR-CPCA-Applied (SMR-CPCA-A).

An advantage of the SMR-CPCA-A compared to the SMR-CPCA-C is its trivial adoption tothe already-introduced Bloom filter tree: it follows the same comparison method (hamming dis-tance), tree construction and traversal as the Bloom filter template. Most of the improvementsby [DRB17] (see section 6.3) are also applicable. Therefore, refer to section 6.2.2 for furtherdetails on the tree construction with the SMR-CPCA-A. The SMR-CPCA-A, compared to theSMR-CPCA-C, loses the ability of merging the auxiliary data for the fractional hamming dis-tance which limits the comparison to the stable bits set in booth templates. Thus the hammingdistance is used for the comparison which compares all set bits, even if some of them are con-sidered unstable in one of the two templates.

59

The SMR-CPCA-C requires a different comparison method (recall section 5.6.2; fractionalhamming distance) and thus requires a tree with nodes that carry auxiliary data. Furthermore,a strategy is needed to handle the two matrices. In the Bloom filter approach, the root node isa union of its two child nodes, both of which are unions of their child nodes, etc. Again, observefigure 7.3 and table 7.1, by taking the union of e.g. the sign-bit of 4 random binary SMR-CPCAthe nodes sign-bit matrix would be already set by 90%. The mask-bit for a union of 4 randombinary SMR-CPCA is also highly populated. Therefore, trivially taking the union of both willnot allow for identification trees containing more than 8 (4 per root children) templates withoutsignificantly losing recognition performance.This problem with the SMR-CPCA-C leads to an artificial third method: replacing the sign-bitof the SMR-CPCA-C with the applied-bit of the SMR-CPCA-A (and keeping the mask-bit)could help to increase the number of templates for the SMR-CPCA-C-Tree and is further calledSMR-CPCA-Mixed (SMR-CPCA-M).

7.1.5 Tree Traversal

The basic tree traversal logic of the CPCA-Tree also follows the same scheme as the Bloomfilter tree. For the SMR-CPCA-A variant, the same comparison method and intelligent traversalstrategies implemented in the Bloom filter tree can be used. In a SMR-CPCA-C- or SMR-CPCA-M-Tree, the comparison method has to be adapted to the sign/mask-scheme by usingthe FHD (see section 5.6.2). Since the FHD yields the same output as the original Bloomfilter tree comparison method (a similarity value), all intelligent traversal strategies are alsoapplicable.

7.1.6 Configuration

On top of the CPCA configuration, there are two variables to be considered for the CPCA-Tree system: SMR-CPCA height (MCP CA) and tree-count (T ). Each of them can be adjustedconsidering the following limitations.

MCP CA The height of the SMR-CPCA templates depends on the training data and the fre-quency (λmin and λmax, section 5.2) of both the training data and model data. Further,the MCP CA is limited by 1 ≤ MCP CA ≤ M , since MCP CA = 0 results in an zero-sizedtemplate and increasing the template (MCP CA > M) is not possible. A large MCP CA

results in an increased template size (thus increased computational workload), and aninsufficient MCP CA drops some necessary information. It has to be noted that MCP CA

has an artificial threshold defined by the mask-bit of each SMR-CPCA. Again, observethe sign-bits shown in figure 7.3: in this figure, the MCP CA = 24 could be further re-duced, but increasing the MCP CA would not increase the recognition performance sincethe mask-bits are not set2 for rows beyond MCP CA = 24.

2Compare figure 5.4f

60

T The behaviour of T for the CPCA-Tree equals T for the Bloom filter tree system and followsthe same limitations: too many trees invalidates the workload reduction introduced bythe system, while too few trees renders the system useless as a worst case.

The three different template types described in section 7.1.4 aren’t considered a configur-ation property, instead they are implementation variants that are benchmarked in the experi-ments.

7.2 Template protection propertiesThe proposed system purely relies on the SMR-CPCA representation. Like mentioned in sec-tion 2.2.4, a biometric system and their templates should address several privacy protectionrequirements. This section inspects the three requirements unlinkability, renewability and irre-versibility for the SMR-CPCA.

7.2.1 Unlinkability

The SMR-CPCA template depends on the training set of the PCA used for the CPCA. It cansafely be stated that two different databases will use two different CPCA training sets.To analyse the unlinkability properties of the CPCA approach, the metrics Dsys

↔ and D↔(s)proposed by Gomez-Barrero et al. [GBRG+16] are used. Dsys

↔ describes the overall systemlinkability and D↔(s) describes the linkability of a single similarity score both measures yielda value on a scale from 0 (fully unlinkable) to 1 (fully linkable). For the following, reasoningDsys

↔ is rated as listed in equation 7.1.

Linkability =

unlinkable if 0 ≤ Dsys↔ < 0.15

semi-unlinkable if 0.15 ≤ Dsys↔ < 0.3

semi-linkable if 0.3 ≤ Dsys↔ < 0.6

linkable if 0.6 ≤ Dsys↔ ≤ 1

(7.1)

Training two CPCA with disjoint training sets (for two different databases) yields two differ-ent transformations (referred to as CPCAA and CPCAB). For this unlinkability experiment,CPCAA is used to generate SMR-CPCAA of SMR1 from instance I1 and CPCAB is usedto generate SMR-CPCAB of SMR1 from instance I1. In other words: same instance, samepalm vein, transformed using two differently trained CPCA. As reference, non-mated templatesare generated with CPCAB to test whether the scores between the mated SMR-CPCAA andSMR-CPCAB are distinguishable from the scores of the non-mated templates. Two differentscores are obtained during the experiment:

Mated scores: scores obtained from the comparison of two templates, extracted from samplesobtained from a single instance I1 and two different transformations CPCAA and CPCAB.

61

0.0 0.1 0.2 0.3 0.4Score

0

2

4

6

8

Prob

abilit

y De

nsity

T ainigset Ove lap = 0%

MatedNon-MatedD% (s)

0.0

0.2

0.4

0.6

0.8

1.0

D%(s

)

(a) 0%→ Dsys↔ = 0.34

0.0 0.1 0.2 0.3 0.4 0.5 0.6Score

0

2

4

6

8

Probability De

nsity

Trainigset Overlap = 10%

MatedNon-MatedD↔ (s)

0.0

0.2

0.4

0.6

0.8

1.0

D↔(s)

(b) 10%→ Dsys↔ = 0.38

0.0 0.1 0.2 0.3 0.4Score

0

2

4

6

8

10

Prob

abilit

y De

nsity

T ainigset Ove lap = 30%

MatedNon-MatedD% (s)

0.0

0.2

0.4

0.6

0.8

1.0

D%(s

)

(c) 30%→ Dsys↔ = 0.48

0.0 0.1 0.2 0.3 0.4 0.5 0.6Score

0

1

2

3

4

5

Probability De

nsity

Trainigset Overlap = 60%

MatedNon-MatedD↔ (s)

0.0

0.2

0.4

0.6

0.8

1.0

D↔(s)

(d) 60%→ Dsys↔ = 0.59

Figure 7.5: Unlinkability Dsys↔ plots for same images at (a) 0%, (b) 10%, (c) 30% and (d) 60% CPCA

training set overlap.

Non-Mated scores: scores obtained from the comparison of two templates, extracted fromsamples obtained from two different instances I1 and I2 and two different transformationsCPCAA and CPCAB

This simulates the scenario when an attacker received two protected templates enrolled in twoapplications (A and B), and tries to decide whether the were extracted from the same instance,e.g. when he tries to track the subject in different databases.However, it is possible that even for different databases there are some small intersections inthe training dataset. To cover this scenario figure 7.5 includes the plots (S = 400) for a trainingset overlap of 0%, 10% and 30% with an additional bad-case overlap of 60%. For 0% (7.5a) and10% (7.5b) overlap a Dsys

↔ < 0.4 is achieved and is already considered semi-, but not fully-,linkable. As could be expected with a strong overlap in the training sets, the templates aremore linkable. In particular, a Dsys

↔ above 0.5 (semi linkable, nearly linkable) is reached withan overlap of 60% (7.5d).

62

There is one additional hurdle for an attacker that is not covered in the previous analysis.Two different databases are usually generated by two different image capturing subsystems,feature extraction subsystems and different images of one subject. When analysing Dsys

↔ for

0.0 0.1 0.2 0.3 0.4Score

0

2

4

6

8

Prob

abilit

y De

nsity

T ainigset Ove lap = 0%

MatedNon-MatedD% (s)

0.0

0.2

0.4

0.6

0.8

1.0

D%(s

)

(a) 0%→ Dsys↔ = 0.11

0.0 0.1 0.2 0.3 0.4Score

0

2

4

6

8

10

Prob

abilit

y De

nsity

T ainigset Ove lap = 30%

MatedNon-MatedD% (s)

0.0

0.2

0.4

0.6

0.8

1.0

D%(s

)

(b) 30%→ Dsys↔ = 0.23

0.0 0.1 0.2 0.3 0.4 0.5 0.6Score

0

1

2

3

4

5

Probability De

nsity

Trainigset Overlap = 60%

MatedNon-MatedD↔ (s)

0.0

0.2

0.4

0.6

0.8

1.0

D↔(s)

(c) 60%→ Dsys↔ = 0.24

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7Score

0

1

2

3

4

5

Probability De

nsity

Trainigset Overlap = 90%

Ma%edNon-Ma%edD↔ (s)

0.0

0.2

0.4

0.6

0.8

1.0

D((s)

(d) 90%→ Dsys↔ = 0.57

Figure 7.6: Unlinkability Dsys↔ plots for different images of the same instance at (a) 0%, (b) 30%, (c)

60% and (d) 90% CPCA training set overlap.

different images but the same image capturing and feature extraction subsystems, the achievedvalues for 0% (7.6a) and 30% (7.6b) are both below 0.3, and thus are considered semi-unlinkable.Even for an overlap of 60% Dsys

↔ = 0.24 (7.6c, semi unlinkable) and for an overlap up to90% Dsys

↔ = 0.57 (7.6d, semi linkable) is achieved. It is safe to assume that with differentSMR configurations in other feature extraction subsystems, Dsys

↔ would further decrease andstrengthen the unlinkability.

63

7.2.2 Renewability

The renewability of SMR-CPCA templates can be reasoned the same way as the unlinkability:the SMR-CPCA template depends on the training set of the PCA used for the CPCA. Therefore,it is possible to renew SMR-CPCA templates in certain boundaries.Observe the similarity-score histogram in figure 7.7. Training two CPCA (refered to as old

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

5 · 10−2

0.1

0.15

0.2

Score

Den

sity

Mated Templates

Non-Mated Templates

Figure 7.7: Histogram of similarity scores for mated old and renewed SMR-CPCA templates andnon-mated old and renewed SMR-CPCA templates.

and renewed) with disjoint training sets yields two different templates: when comparing twoSMR-CPCA of one instance generated with the old and the renewed CPCA, the score cannotbe distinguished between scores generated by comparing the old SMR-CPCA with renewedSMR-CPCA of other instances.

Another hint at the renewability capability of the SMR-CPCA can be derived from theanalysis of the unlinkability in the previous section. Observe the unlinkability for e.g. 30%and 60% training set overlap in figure 7.6. When acquiring new images and not using a fullydisjoint training set for the new CPCA to renew the templates (e.g. after a data leakage), theunlinkability is still given to some extent. By contrast, if the new template is not linkable tothe old template, the renewability is also given.

7.2.3 Irreversibility

The SMR itself can be stated irreversible3 (it cannot be transferred back to its minutiae), al-though a discourse to prove the irreversibility of an SMR template is beyond the scope of thisthesis. However, the SMR can be defined as a raw template, since the renewability is not givenand the unlinkability is only given with e.g. permutations. Therefore, it is necessary to analysethe irreversibility properties of the CPCA transformation, to determine if the SMR-CPCA canbe stated irreversible.

3[XV10c] hints at the irreversibility of the SMR but never actually states nor proves it.

64

In [XVKA09], the CPCA is given by

XCPCA = UTZX (7.2)

where UZ is the transformation matrix used to project the SMR spectrum X to its SMR-CPCA XCPCA. By only keeping the first MCP CA columns of UZ, UZ is received, thus yieldingthe MCP CA×N sized SMR-CPCA. The CPCA can trivially reversed4 by

XREV = UZXCPCA. (7.3)

This introduces a security risk since the reversibility renders both unlinkability and renew-ability useless. Therefore it is necessary to treat UZ (and UZ) as classified key matrix for theCPCA system that should not be saved as plain data.

It has to be noted that the trivial reversibility is aggravated for the attacker in a real-worldscenario, since he has to undertake additional steps to sufficiently reconstruct a SMR template.The SMR-CPCA templates should be stored in their normalised form. Since the normalisationis applied after the CPCA transformation, the dynamic range of the CPCA matrix multiplic-ation (usually [−3, 3]) shrinks during the normalisation process (normalised to [−1, 1]). Whenreversing the matrix multiplication - which would normally transform the pre-normalised SMR-CPCA dynamic range to the normalised SMR - the resulting SMR is also reduced in its dynamicrange. Comparing the reversed SMR with the original SMR using the direct comparison methodtherefore yields an average similarity score of ∼0.4. However, the original structure of the SMRis retained during the transformation process. Therefore, applying the normalisation on thereconstructed SMR yields SMR that achieve an average similarity score of ∼0.998 when com-pared to the original SMR. Reconstruction of normalised SMR-CPCA templates is not perfectbut more than sufficient for tracking or other purposes.

Facing the general reversibility, a possible scenario is that an attacker was able to receivethe SMR-CPCA template but could not receive UZ since it is strongly encrypted. The attackercan train an approximated UZ (referred to as U′

Z) by training a CPCA with random SMRtemplates that feature the same characteristics (M , N , SML/SMC) and approximate configur-ations (λmin, λmin, σ).Again, the unlinkablity measure Dsys

↔ from [GBRG+16] is used to analyse the scenario. Ob-serve figure 7.8, the linkability for the reconstructed SMR (here SML) while using U′

Z by 0%CPCA training set overlap is with 0.23 still considered semi-unlinkable. In scenarios where theattacker was able to acquire templates used in the training set of the CPCA used to createUZ, he is able to generate a U′

Z with an overlap. For U′Z with 50% overlap, the SMR-CPCA

4The error introduced in double precision implementations is hardly detectable. For single precision imple-mentations, the error is noticeable but influences a comparison between XREV and X by less than 0.5% inexperiments.

65

0.00 0.05 0.10 0.15 0.20 0.25Score

0

2

4

6

8

10

Prob

abilit

y De

nsity

T ainigset Ove lap = 0%

MatedNon-MatedD% (s)

0.0

0.2

0.4

0.6

0.8

1.0

D%(s

)

(a) 0%→ Dsys↔ = 0.23

0.00 0.05 0.10 0.15 0.20 0.25Score

0

2

4

6

8

10

Prob

abilit

y De

nsity

T ainigset Ove lap = 20%

MatedNon-MatedD% (s)

0.0

0.2

0.4

0.6

0.8

1.0

D%(s

)

(b) 20%→ Dsys↔ = 0.16

0.00 0.05 0.10 0.15 0.20 0.25Score

0

2

4

6

8

10

Prob

abilit

y De

nsity

T ainigset Ove lap = 50%

MatedNon-MatedD% (s)

0.0

0.2

0.4

0.6

0.8

1.0

D%(s

)

(c) 50%→ Dsys↔ = 0.24

0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35Score

0

2

4

6

8

10

Probability De

nsity

Trainigset Overlap = 80%

MatedNon-MatedD↔ (s)

0.0

0.2

0.4

0.6

0.8

1.0

D↔(s)

(d) 80%→ Dsys↔ = 0.59

Figure 7.8: Unlinkability Dsys↔ plots between reconstructed SML and original SML at (a) 0%, (b) 20%,

(c) 50% and (d) 80% CPCA training set overlap.

is still considered semi-unlinkable. When reaching up to 80% overlap, Dsys↔ = 0.6 is considered

semi-linkable. Exceeding 80% overlap quickly rises Dsys↔ and 100% overlap results in Dsys

↔ = 1.Note the linkability of 0.16 for 20% overlap is smaller than the linkability of 0% overlap. Thisbehaviour5 was observable in all intervals of this experiment6.

These linkability scores were generated using training sets and templates with equal SMRcharacteristics and configurations, and therefore they represent perfect conditions for an at-tacker since he is using a perfect copy of the system. Again, it is safe to assume that thelinkability decreases when different characteristics and configurations are used in the training

5Between 0% and 50% overlap, one or two linkability scores peaked (where smaller) compared to linkabilityscores generated with less overlap.

6This author is not sure about the origin of this behaviour. One possible explanation would be that someSMR have a higher impact on the CPCA training, thus when these are included in both training sets, U′

Z andUZ become more equal; when the overlap increases, the other SMR even the impact of the high impact SMR.

66

sets and templates.

7.3 ConclusionsThe CPCA-Tree is a novel workload reduction approach exclusively for SMR-based biometricsystems. It adopts the same strategy and mechanism of the already-proven Bloom filter-indexingapproach [RBB13, RBBB15, DRB17].

In the first section of the chapter, the basics of the system were outlined. The templatetransformation uses the CPCA projection to reduce the full SMR feature vector to a on av-erage 80% smaller representation. To construct a single CPCA-Tree, three different methodsfor representing a template are presented. Which method yields the best retrieval performancewill be explored in the experiments. For the tree traversal, all state-of-the-art strategies from[DRB17] are also applicable. Merely the comparison method is adapted to the three differenttemplate methods. Finally, the CPCA-Tree features a configuration comparable to the Bloomfilter-indexing configuration.

The second section analyses the template protection properties unlinkability, renewabilityand irreversibility. All three properties were evaluated using the unlinkablity measure Dsys

↔ from[GBRG+16]. While the unlinkability and renewability are given without prerequisites, the irre-versability is only given if the transformation matrix UZ of the CPCA approach is unreachablefor an attacker (e.g. with strong encryption). When the unlinkability, renewability and irre-versibility reached with the CPCA transformation are not sufficient, permutation approacheslike the presented method in [GBRG+16] are also applicable.

The CPCA-Tree indexing approach will be used as a contender to the already-proven Bloomfilter-indexing approach in the workload reduction experiments.

67

Chapter 8

Experimental Setup

The following sections describe the palm vein data used during this project for experiments, itspreparation and a description of how the experiments were conducted.

8.1 DatasetsWhile there are many publicly-available vascular datasets, only few of them contain palm veinimages. This situation is aggravated by the fact that most palm vein images are captured usinga self-designed apparatus resulting in noisy and poor-quality images. However, for use duringthe conceptual and experimental phases of this project, three distinct datasets were selected.The datasets are briefly described below, example images are presented in figure 8.1 and alldataset values are summarized in table 8.1.

PolyU multispectral palmprint Database (PolyU) [pol] At the time of writing, the PolyUdataset is the largest publicly-available palm-print dataset containing NIR palm-printimages usable for palm vein recognition known to the author. It comprises images fromtwo sessions in an average of 9 days difference of 250 subjects with 6 images per handand session, resulting in 6 000 images. The images have a pre-defined and stable ROI.All images have a very low-quality variance and are all equally illuminated. Since thePolyU dataset aims for palm-print recognitionm, it features a high amount of skin texture,which interferes with the vein detection and makes it a challenging dataset for the featureextraction pipeline.

CASIA-MS-PalmprintV1 (CASIA) [cas] Another palm-print targeting dataset featuringNIR palm-print images usable for palm vein recognition. The dataset is built in twosessions, with a difference of more then one month, from 100 subjects with 3 images perhand (i.e. biometric instance) and session, resulting in 1 200 images. All images picturethe full geometry of differently-shaped hands without any guidance, thus makes it aprefect dataset for experiments for guidance-less ROI detection. Unfortunately, the imagecapturing device is a multispectral device, thus missing any visible light filters resulting

68

(again) in images with a high amount of skin texture. Furthermore, the illuminationvariance is very high, since the capturing device used specular LEDs. Therefore, smallhand movements result in large illumination (shadows, hotspots) variance. Some handswere bent or crooked during the capturing process, aggravating the feature extractionor resulting in failure-to-acquire events. In summary, it is the most challenging dataset,which is expected to yield very low performance.

PUT Vein Database (PUT) [KK11a] Additional small dataset with three sessions, with adifference of at least one week, including 50 subjects with 4 images per hand and session,resulting in 1 200 images with a pre-defined ROI. The ROI suffers from a high translationvariance but the images feature almost no skin texture or other noise1.

In the PolyU dataset it is not possible to link the left hand instance and the right hand instanceto one subject. The CASIA and the PUT datasets link left and right hand instances to subjects.Since it isn’t possible to link left and right hand instances to subjects by just viewing at thepalm vein images, it is assumed in the following sections and chapters that no two instacesare from one subject and therefore, every subject is represented by only one instance in thedatasets.

Dataset Instances Images Excluded Instances§ Excluded Images§ Resolution ROI Quality

PolyU 500 6 000 6 50 128×128 px 128×128 px ModerateCASIA 200 1 200 28 30 768×576 px max. 128×128 px LowPUT 100 1 200 0 0 512×384 px* 272×176 px† Moderate‡

* Downscaled off-line to reduce resource usage (disk and computation).† Original 340×240 px downscaled by factor 0.8.‡ High image quality but no ROI alignment possible.§ Refer to section 8.3 for a detailed explanation of the exclusion criteria.

Table 8.1: Dataset overview.

(a) PolyU (b) CASIA (c) PUT

Figure 8.1: Example images from each dataset.

1It could be possible that the images are heavily pre-processed by the image-capturing device or some othersubsystem in terms of contrast and colour gradients. All images are differently saturated and coloured, whichhints at some pre-preprocessing steps.

69

8.2 Palm Vein Feature Detection PipelineThe complete biometric processing chain (pipeline) has already been presented in chapter 4.Since the data of the selected datasets features completely different characteristics, the pipelinehas to be adjusted to the three datasets.

CASIA The CASIA dataset uses the full featured pipeline presented in chapter 4.

PolyU Since the PolyU dataset already features a pre-defined ROI, the ROI detection step(section 4.2) is skipped. All following steps are used as presented.

PUT With the already pre-defined ROI of the PUT dataset, again no ROI detection step isneeded. However, the images are further tailored to remove the background, thus removingnoise. Refer to the appended figure A.1 for a visualisation of the ROI used.

8.3 Excluded ImagesUnfortunately, not all images or subjects could be used for the experiments. This section outlinesthe reasoning behind the decisions for excluding images.

8.3.1 Capturing Errors

As described in section 2.1, capturing vascular images is not a trivial task. The four differentfailure-to-acquire types observed through the datasets are outlined below and example images2

are presented in figure 8.2.

No visible veins For some instances, no veins are visible at all. Even with manual imagemanipulation, no (real) veins were extractable. This is mostly due to a poorly-chosenNIR wavelength and when the veins are too deep inside the skin. In all cases, this affectsthe whole subject, resulting in an exclusion of all images of that instance.

Missing/additional veins In some occasions, the pulses can make some veins visible, if theimage was taken at the exact time of the heartbeat that are not visible between heartbeats.Several images were found where the minutiae count fluctuated more than 50% due tothe pulses. To remove as few images as possible, the images with a high minutiae count(due to the heartbeat) were identified and removed as more images between heartbeatsthan images on a heartbeat are expected.

Misplaced hand Only affecting the CASIA dataset, some hands were shifted or tilted duringduring the image-capturing process. This heavily warps the venous network and resultsin false negative identifications.

2Missing/additional veins are difficult to illustrate using only one example image. The full illustration wouldexceed the scope of this section.

70

Bend hand Again only affecting the CASIA dataset, some hands were bent or crooked duringimage capturing. This mostly created skin folds and shadows that were detected as veinswhile covering the real veins. Consequently, bent hands were falsely mated rather thanmated with their enrolled sample.

Inhomogeneous illumination Again only affecting the CASIA dataset, the inhomogeneousillumination created hotspots and shadows depending on the positioning of the hand. Insome images, the shadows and hotspots displaced the real veins, resulting in falsely-matedtemplates.

(a) No visible veins (b) Misplaced hand

(c) Bend hand (d) Inhomogeneous illumination

Figure 8.2: Example images for the different failure-to-acquire event types.

Refer to table 8.1 for a summary ofhow many images are excluded for each dataset.Note: no images of the CASIA dataset were removed due to quality issues. Therefore, the datasetcontains all data from misplaced, bent or inhomogeneous illuminated hands. It is expected thatthis will significantly reduce the recognition performance on this dataset.

8.3.2 Preprocessing Failures

Preprocessing failures can occur as early as the ROI detection. The ROI is the most vulner-able and critical preprocessing step. A misplaced ROI propagates further down the processingchain, is difficult to automatically detect and impairs the overall system performance. Figure8.3 illustrates the three common ROI detection errors and is explained below. The presentedROI detection approach in section 4.2 tries to determine the rotation of the hand and which

71

(a) (b) (c)

Figure 8.3: Example images for reasons of a misplaced ROI: (a) wrongly interpreted rotation; (b) onegap detected as two gaps; (c) wrongly associated gaps.

side of the hand is captured. In some cases, this step fails and returns a wrong association ofthe finger gaps, as illustrated in figure 8.3a. When the segmentation step fails to extract a cleancontour, the finger gap finding step detects two gaps at one location. This usually yields anROI smaller than 10px, illustrated in figure 8.3b, that is misplaced as well. Lastly, the errorillustrated in figure 8.3c is again a result of a falsely detected hand in terms of left and righthand. In this example, the subject has presented the left hand. The ROI detection falsely clas-sifies it as a right hand and the gap association step associated the thumb-index-finger gap asthe ring-little-finger gap, etc. Therefore, the determined ROI is too large and thus exceeds thehand perimeters.

Aside from the error illustrated in figure 8.3b, these errors are notoriously difficult to detect.Through a visual inspection, the worst and most obvious failures were found. From the CASIAdataset, 30 images (2.5%) had to be removed to avoid including templates from invalid ROIsin the experiments.Again, refer to table 8.1 for a summary of how many images are excluded for each dataset.

8.4 ExperimentsThe dataset has been split into four groups: enrolled, genuine, impostor and training (forthe CPCA feature reduction). Table 8.2 shows the number of templates in each group. Theconducted experiments are listed below.

8.4.1 Conducted experiments

This section outlines the experiments conducted as practical work for this thesis.

MHD-Baseline The basic implementation, comparing palm vein probes based on minutiaewith the modified Hausdorf Distance (MHD) [DJ94]. In an identification, scenario the

72

Dataset Enrolled templates Genuine samples Impostor samples Training samples

PolyU 256 (256) 2 816 (256) 2 487 (238) 100 (6)*

CASIA 99 (99) 476 (99) 468 (72) 100 (28)†

PUT 58 (58) 456 (58) 456 (38) 48 (4)* With added images previously excluded.† Only ∼4 images (2 per session) used per subject.

Table 8.2: Dataset partitioning overview, subjects in parentheses.

database is searched exhaustively, e.g. every query template (probe) is compared withevery enrolled template (reference).

Verification Each template of every subject is compared with another template of thesame subject and the score is checked against a threshold. Impostor scores are gen-erated by comparing every template of one subject with all templates of anothersubjects. A false match is counted if a score of one non-mated comparison trial ex-ceeds the decision threshold, whereas a true match is counted if a score of a matedcomparison trial exceeds the decision threshold, and a false non match is counted,if a score of one mated comparison trial does not exceed the decision threshold.

Identification One reference template is enrolled for each subject. All remaining probetemplates are compared against the enrolled references. A false positive identificationis counted if the highest score of one probe exceeds the threshold on a referencetemplate of a wrong subject (impostor score), whereas a true positive identificationis counted if the highest score of one probe exceeds the threshold on an genuinereference template, and a false negative identification is counted if the highest scoreof one probe does not exceed the threshold.

SMR The same procedure as the baseline experiments, but with the SMR and direct SMRcomparison. Since the main focus of this project is identification in large-scale databases,the identification experiments for the SMR are used to find the optimal representationand SMR settings for following experiments. These experiments represent the baselinefor the workload-reduction approaches in the following experiments.

Identification As in the baseline identification experiment above, but with the SMRand direct SMR comparison. This setup is used to find the optimal representationand SMR settings.

Verification Supplemental experiment to the identification experiments. Procedure asdescribed in the baseline verification experiments.

Workload reduction by feature reduction After determining of the best-performing SMRconfiguration as the performance and workload baseline, the first workload reduction

73

approaches for palm vein SMR are applied and evaluated in identification and verificationscenarios as above.

CPCA Repetition of the above verification experiments, but with CPCA feature reduc-tion.

Binary SMR The same experiments (for both original SMR and SMR-CPCA, verifica-tion and identification) as above are repeated with the binary representations of theSMR. It holds special interest whether the binary recognition performance benefitsfrom other SMR settings than the (original) double-precision SMR.

Bloom filter The second workload-reduction experiments with the system described in chapter6.

Identification and workload reduction - Binary SMR-CPCA With one templateper subject (references), the Bloom filter trees are constructed. Genuine and impostorscores are obtained as described in the baseline implementation. All results achievedin the means of biometric performance are analysed with respect to the achievedworkload reduction on top of the results achieved in the workload reduction byfeature reduction section.

Identification and workload reduction - Binary SMR An additional experimentis run the same way as the above experiment to test whether the larger binaryrepresentation achieves a higher biometric performance under equal workload re-duction achievements.

Verification Completing the Bloom filter experiments, the verification capabilities forboth binary SMR-CPCA and binary SMR Bloom filter templates are tested.

CPCA-Tree The third workload-reduction experiments with the system described in chapter7. These experiments follow the same procedure as the experiments for the Bloom filter,without the verification experiments since the templates are the same as in the workloadreduction by feature reduction binary SMR-CPCA section.

Full vascular pattern SMR Finally, additional experiments are run to test whether theSMR can be enhanced with the full vascular pattern to increase the biometric perform-ance.

8.4.2 Experiment vocabulary

Alongside plots and tables, many labels employing shortcuts are used. The following table 8.3lists all labels and outlines their meaning.

74

Label Reference Description

MHD-Baseline 9.1 Biometric performance baseline yielded by the MHDidentification and verification experiments

PSML-Baseline 9.2.4 Biometric performance baseline yielded by the naïve SMRidentification and verification experiments for the PolyU dataset

PSML-CPCA 9.3.2 Biometric performance (or the experiment) yielded byapplying the naïve CPCA transformation to the PSML-Baseline

BinaryPSML

9.3.2 Biometric performance (or the experiment) yielded by binarisationof the naïve PSML-Baseline

BinaryPSML-CPCA

9.3.2 Biometric performance (or the experiment) yielded by binarisationof the PSML reduced with the CPCA feature-reduction (PSML-CPCA)

SM* Tuned 9.2.1, 9.2.2,9.2.3

SMR types tuned in the means of λmax

Bf 9.4.2 Bloom filter-indexing experiment with binary PSML templates

Bf ABC 9.4.2 Bloom filter-indexing experiment with binary PSML templatesand an additional binary PSML comparison

Bf CPCA 9.4.1 Bloom filter-indexing experiment with binary PSML-CPCA templates

Bf CPCA ABC 9.4.1 Bloom filter-indexing experiment with binary PSML-CPCA templatesand an additional binary PSML-CPCA comparison (ABC)

Bf CPCA ARC 9.4.1 Bloom filter-indexing experiment with binary PSML-CPCA templatesand an additional real-valued PSML-CPCA comparison (ARC)

PSML-CPCA-A 7.1.4, 9.5.1 SMR-CPCA-A (SCA) templates build from PSML-CPCA-A templatesPSML-CPCA-C 7.1.4, 9.5.1 SMR-CPCA-C (SCC) templates build from PSML-CPCA-C templatesPSML-CPCA-M 7.1.4, 9.5.1 SMR-CPCA-M (SCM) templates build from PSML-CPCA-M templatesC-T SCA 9.5.1 CPCA-Tree indexing experiment with PSML-CPCA-A templatesC-T SCC 9.5.1 CPCA-Tree indexing experiment with PSML-CPCA-C templatesC-T SCM 9.5.1 CPCA-Tree indexing experiment with PSML-CPCA-M templates

C-T SC* ARC 9.5.5 CPCA-Tree indexing experiment with PSML-CPCA-A/C/M templatesand an additional real-valued PSML-CPCA comparison

C-T SC* Masking 9.5.4 CPCA-Tree indexing experiment with PSML-CPCA-A/C/M templatesand masking out most common set bits

ω256 8.5 Workload ω for S = 256𝟋B 8.5 Workload fraction 𝟋 compared to the (naïve) PSML-Baseline𝟋R 8.5 Workload fraction 𝟋 compared to the (naïve) binary PSML-CPCA

Table 8.3: Labels used in plots and tables.

8.5 Workload metricAside from the performance reporting metrics specified by the ISO/IEC 19795-1 [iso06], thesystems workload and system workload reduction metrics introduced by [DRB17] are used tocompare the systems workload for the different approaches and expansion stages.The authors of [DRB17] present six requirements for a full-featured workload and workload-reduction reporting:

R1 The baseline workload must be explicitly stated.

75

R2 The baseline biometric performance of a state-of-the-art algorithm on the dataset usedmust be explicitly stated in a manner described in the ISO/IEC standard.

R3 The workload of the proposed scheme is to be stated in the manner described in R1.

R4 The biometric performance of the proposed scheme must be reported according to theISO/IEC standard.

R5 The additional costs and benefits of the proposed scheme should be listed.

R6 The total workload for both the baseline and the proposed system is to be computed usingequation 8.1.

The workload itself is expressed asω = S ∗ p ∗ τ (8.1)

where S is the number of enrolled subjects, p the penetration rate (measure of the averagenumber of pre-selected templates as a fraction of the total number of templates, as defined in[iso06]) and τ the cost of a single comparison (i.e. the number of compared bits). The measureω is not feasible to describe the workload for the chosen baseline since the minutiae featureset yields a varying τ and the comparison cost of the modified hausdorf distance with varyingreference points is difficult to describe comparable to bit comparisons. Therefore, the workloadof a naïve SMR system is used as a baseline. The workload reduction is stated as a fraction (𝟋)of the baseline workload as proposed by requirement 6.

8.6 SummaryThis chapter has presented the details of the experimental setup for this project. In the first sec-tion, the datasets used and their properties are outlined. The following two sections describedthe feature detection pipeline and its errors. Next, the roadmap of the performed experimentswas presented. The results of these experiments were presented and discussed in the followingchapters. Finally, the workload metric used by [DRB17] was introduced.

The three chosen datasets do not feature optimal raw data. However, they were chosen dueto the lack of better palm vein datasets. The PolyU dataset features the best characteristics,albeit it is not flawless (instances without visible veins). To test the ROI detection and receive asecond test database, the CASIA dataset is used. Unfortunately, all contained palm vein imagescontain a high amount of skin texture (noise) and some subjects bent, misplaced or shiftedtheir hands. This dataset is expected to yield the worst results. With the PUT dataset, the onlypurpose built palm vein dataset is introduced. While it features a high image quality, it is notpossible to automatically align the ROI since there are no reference points of the hands visible,and therefore the translation invariance features of the SMR has to compensate for that.

76

Chapter 9

Results

The experimental results of this project are presented in this chapter.

9.1 BaselineThe baseline results are obtained using the MHD for the extracted palm vein minutiae. Theseresults are necessary to evaluate the results of the subsequent experiments and were used tofind the base settings of the feature extraction pipeline for each dataset. All experiments whererun with inter-session data.

9.1.1 Score Distribution

Looking upon the similarity score distribution of the datasets provides a first hindsight into thequality of the biometric data. Figure 9.1 presents the best-achieved MHD similarity score dis-tributions for the three chosen datasets with feature extraction pipelines configured as statedin table 9.1. Observe h (NL-means) for all three datasets: the experiments have shown that

max.-cur. NL-means NL-diffusion

Dataset σ h Sizetemplate† Sizesearch

† k *

PolyU 8 7 7 23 1CASIA 7 7 7 23 2PUT 7 7 7 23 1* Out of 2 non-linear-scale-space iterations with 2 sub-iterations.† Same for all three datasets: ROI size is fixed to fit these sizes.

Table 9.1: Configuration of the minutiae feature extraction pipeline: σ - kernel size of the maximumcurvature algorithm; h - decay parameter of the Euclidean distance in the NL-means; Size - templatewindow and search kernel size of the NL-means algorithm; k - amount of diffusion iterations in theNL-diffusion algorithm.

77

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

Score

Den

sity

Mated Scores

Non-mated Scores

(a) PolyU

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

Score

Den

sity

Mated Scores

Non-mated Scores

(b) CASIA

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

Score

Den

sity

Mated Scores

Non-mated Scores

(c) PUT

Figure 9.1: MHD similarity score distribution (normalised to x = 1.0) for the three datasets.

h > 7 does not increase the recognition performance in any means1. In some cases, a h > 7reduces the recognition performance because contrast resulting from noise (i.e. skin-texture) isalso increased. Further observations showed that the maximum-curvature algorithm does notnecessarily needs very high contrast to properly extract the veins, and thus reducing h to 7does not impair the maximum-curvature performance.

1This contradicts the observation stated in section 4.3.1.

78

From examining figure 9.1, the following two observations can be stated:

• The score distribution for the PUT dataset is very bad. The nearly full overlap of impostorand genuine comparison trials yield no clear threshold, thus it is impossible to distinguishbetween impostor and genuine probes. This is expected for the PUT dataset since thedataset features many characteristics (rotations, translations, spurious minutia) that theMHD is not capable of compensating.

• The PolyU dataset impostor and genuine score overlap is very small. A threshold of around0.65 is capable of distinguishing between impostor and genuine comparison trials with avery small error compared to the PUT dataset.

• As expected in section 8.3.1, the bad quality data contained in the CASIA dataset shiftsthe genuine scores to the left2.

Basing on these observations and the three score distributions, it is expected that the PolyUdataset will yield the best biometric performance, followed by the CASIA dataset with a distantperformance for the PUT dataset.

9.1.2 EER and ROC

The EER is a commonly-used metric for reporting the performance of a biometric verificationsystem. It is defined by the [iso06] as the point at which the FMR and FNMR are equal. The

Dataset EER

PolyU 5.4%CASIA 30.5%PUT 47.3%

Table 9.2: EER for the MHD comparison of the PolyU, CASIA and PUT datasets.

conclusions derived from figure 9.1 in the previous section can be validated using the EERpresented by table 9.2. With an EER of 5.4%, the MHD achieves an moderate verificationperformance for the PolyU dataset. As expected, the performance of the CASIA dataset is belowthe PolyU dataset. However, the EER of 30.5% for the CASIA dataset confirms the MHD is notable to achieve distinguishable scores of impostor and genuine probes with the bad quality dataof the dataset. Also confirming the expectations for the PUT dataset with an EER of 47.3%,the MHD fails to yield any distinguishable scores for heavily translated or rotated data.

2It can be expected that the CASIA dataset would perform similar to the PolyU dataset if the poor qualitydata were excluded.

79

The relevant metrics for biometric identification systems are the True Positive IdentificationRate (TPIR) and FPIR, both defined by the [iso06] standard. By moving the decision thresholdfor the biometric identification system, these metrics can be plotted as a Receiver OperatingCharacteristic Curve (ROC) curve for fixed database sizes. Besides the ROC curve, three ad-ditional biometric performance points are used: TP0.5, TP0.1 and TP0, which correspond tothe TPIR at 0.5%, 0.1% and 0% FPIR, respectively. These will be used where single-valuebiometric performance indicators are feasible.The ROC curves for the MHD comparison are plotted in figure 9.2 with TP0.5, TP0.1 and

10−4 10−3 10−2 10−1 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate PolyU

CASIAPUT

Figure 9.2: ROC curves for the MHD baseline system

Dataset TP0.5 TP0.1 TP0

PolyU 63.6% 59.7% 49.8%CASIA 14.1% -* 12.7%PUT 4.4% -* 4.1%* Database to small to determine

TP0.1.

Table 9.3: TP0.5, TP0.1 and TP0 for the MHD comparison of the PolyU, CASIA and PUT datasets.

TP0 listed in table 9.3. As expected facing the results of the verification experiment, the MHDachieves the best biometric performance on the PolyU dataset, while CASIA and PUT are laggingfar behind with an unacceptable TP0. However, with a TP0.1 of 59.7% and a TP0 of 49.8%,the MHD yield results far from an acceptable performance.

9.2 Spectral Minutiae RepresentationThe initial SMR approach in [XVK+08] targeted fingerprints and their minutiae for biometricidentification and verification. Image capturing of fingerprint samples is a much cleaner processthen the image capturing of the vascular pattern. In fingerprint recognition the observed bio-metric characteristics are on the sufrace of the skin, and thus are easily collectable in various

80

ways. However, recall section 2.1, the palm veins are internal to the individual’s body andonly an blurred approximation of the vein pattern is represented in the captured vein images.Thus, even with a very stable feature extraction pipeline, more noise than in fingerprint featureextraction case must expected.All three datasets show different properties. Therefore, it is necessary to run exhaustive exper-iments for all datasets since assumptions and results for one dataset are hardly applicable forthe other datasets.

9.2.1 PolyU dataset

The PolyU dataset achieved the best biometric performance on the MHD baseline experiments.It features the best characteristics among all three datasets: homogeneous illumination, notranslation, consistent noise levels and most poor-quality images have been removed. The in-put minutiae for the experiments are generated using the same feature extraction pipeline andsettings as presented in table 9.1.

In section 5.1, three SMR types are presented. Further, every type can be compared absolute-valued or real-valued and additionally supplied with quality informations about single minutiae.Since the SMO does not yield a better performance than the SML (as stated by the authors in[XVB+09]), only SMC and SML will be used in the experiments. As shown in figure 9.3 for both

10−3 10−2 10−1 100

0.1

0.2

0.3

0.4

0.50.5

0.6

0.7

0.8

0.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

MHD-BaselineSMC (Absolute)SMC (Real)SML (Absolute)SML (Real)QSMC (Absolute)QSMC (Real)QSML (Absolute)QSML (Real)

Figure 9.3: ROC curves of the SML, SMC, QSML and QSMC type for the PolyU dataset.

types the real-valued comparison performs much better than the absolute-valued comparison.The best performance for basic SMR identification approach is achieved by the real-valued SMLwith TP0.1 = 75.1% and TP0 = 72.4% followed by the real-valued QSML with TP0.1 = 77.3%and TP0 = 70.1%. All six experiments were run with the settings recommended in [XVB+09];σ = 0.32, λmin = 0.1 and λmax = 0.6. Note: the quality information extracted by the maximum-curvature algorithm for the quality data enhanced SMR (QSMR) is multiplied by 4 to achievethe plotted results. Refer to appended figure A.2 for a comprehensive plot of gain 2 to 8 forthe QSML representation. Further, no qL data were used for the QSMR experiments: the qL

81

data made little to no difference in the pre-experiments. In theory, the QSML should performbetter than the SML. Refer to the discussion (chapter 10) why the QSML is not able to achievea higher performance in these experiments.

At this point, the real-valued SML, QSML and SMC representation achieve higher biometricperformances than the basic MHD approach.

As mentioned in section 5.8, it is possible to use a minutiae pre-selection to exclude potentialspurious minutiae. Figure 9.4 shows the ROC curves for the five minutiae selection approachesapplied for the real-valued SML, SMC, QSML and QSMC3.Only the SMC and QSMC profit from a minutiae pre-selection, albeit insufficiently to reach

10−3 10−2 10−1 100

0.10.20.30.40.50.50.60.70.80.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhoodqM > 0.1qM > 0.2Bifurc. & qM > 0.1Bifurc. & qM > 0.2

(a) SML

10−3 10−2 10−1 100

0.10.20.30.40.50.50.60.70.80.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhood

(b) QSML

10−3 10−2 10−1 100

0.10.20.30.40.50.50.60.70.80.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhoodqM > 0.1qM > 0.2Bifurc. & qM > 0.1Bifurc. & qM > 0.2

(c) SMC

10−3 10−2 10−1 100

0.10.20.30.40.50.50.60.70.80.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhood

(d) QSMC

Figure 9.4: ROC curves of the real-valued SML, SMC, QSML and QSMC type for the PolyU datasetwith applied minutiae pre-selection for determining a PSMR configuration.

the performance of the SML.After determining the most promising SMR types, the final step is to tune the SMR sampling

to fit the quality of the data (noise level) as mentioned in [XVB+09]. As for the SML, the tuninghas not achieved a better biometric performance in terms of TP0.1 and TP0. However, in termsof the overall performance the tuned sampling does not achieve much better results.

The top three configurations in terms of identification biometric performance for the PolyU3Since the QSML and QSMC already include quality information (see section 5.7), the reliability selection

is skipped.

82

10−3 10−2 10−1 100

0.7

0.8

0.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

SML (Tuned)QSML TunedQSMLSMC (Bifurc.) TunedSMC (Bifurc.)

Figure 9.5: ROC curves of the most promising SMR types, tuned and untuned, for the PolyU dataset.Note: the tuned SML and the untuned SML are the same ROC since the SML achieves the bestbiometric performance at λmax = 0.6.

dataset are summarized in table 9.4. For each of the top three configurations, a verificationexperiment has been run and the EER is also stated.

SMR Ident. Verif.

Type Pre-Selection λmin λmax TP0.5 TP0.1 TP0 EER

SML - 0.1 0.6 78.1% 75.1% 72.4% 3.4%QSML - 0.1 0.64 80.8% 76.2% 70.5% 3.3%PSMC Bifurcations 0.1 0.57 72.2% 69.4% 68.0% 7.1%

Table 9.4: Top three configurations (TP0.1), for a SMR biometric identification system using the PolyUdataset, with their corresponding verification EER, ordered by TP0.

9.2.2 CASIA dataset

The CASIA achieved the second best, but worse, biometric performance on the MHD baselineexperiments. It features the worst characteristics among all three datasets: inhomogeneous illu-mination, inconsistent noise levels, misplaced or tilted hands and no poor-quality image, exceptthose with a falsely-detected ROI, have been removed. Again, the input minutiae for the ex-periments are generated using the same feature extraction pipeline and settings as presentedin table 9.1.

Repeating the same experiments as for the PolyU dataset above, the initial SML, SMC,QSML and QSMC plots are presented in figure 9.6. The real-valued comparison again outper-forms their absolute-valued counterpart. As expected from the MHD results, the results usingSMR are not acceptable either. Unfortunately, the original SMR approaches even perform worsethan the MHD approach.

83

10−3 10−2 10−1 100

0.1

0.2

0.3

0.4

0.5

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

MHD-BaselineSMC (Absolute)SMC (Real)SML (Absolute)SML (Real)QSMC (Absolute)QSMC (Real)QSML (Absolute)QSML (Real)

Figure 9.6: ROC curves of the SML, SMC, QSML and QSMC type for the CASIA dataset.

As already applied to the PolyU dataset, the CASIA dataset might profit from applying vari-ous minutiae pre-selection approaches. Since the real-valued SMR again yield a better basis,the very same minutiae pre-selection approaches as for the PolyU dataset are applied for thePolyU dataset on the real-valued SMR in figure 9.7. Unfortunately, even the various minutiaepre-selection approaches are unable to compensate the bad quality of the CASIA dataset. Onlythe performance of the SMC is improved by removing all endpoint minutiae.

The last outstanding experiment for the CASIA dataset is to tune the λ values for thedifferent, most promising SMR. Since only the SMC and the only-bifurcation-minutiae-PSMCyield results that can be concerned, the plot in figure 9.8 only features the tuning of these types.Finally, table 9.5 summarizes the best configurations as for the CASIA dataset. The performance

SMR Ident. Verif.

Type Pre-Selection λmin λmax TP0.5 TP0 EER

PSMC Bifurcations 0.1 0.60 27.9% 27.6% 16.4%SMC - 0.1 0.45 24.6% 22.8% 17.5%

Table 9.5: Top two configurations (TP0), for a SMR biometric identification system using the CASIAdataset, with their corresponding verification EER.

on the CASIA dataset is astonishingly poor. Compared to the MHD approach, the SMR gainsaround 10% points in the means of TP0 and TP0.5. These poor performances (MHD and SMR)hint that there might be a problem with the feature extraction pipeline.

9.2.3 PUT dataset

Finally, the PUT achieved the worst biometric performance on the MHD baseline experiments.The results are dominated by two good and two bad properties: homogeneous illumination and

84

10−2 10−1 100

0.1

0.2

0.3

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhoodqM > 0.1qM > 0.2Bifurc. & qM > 0.1Bifurc. & qM > 0.2

(a) SML

10−2 10−1 100

0.1

0.2

0.3

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhood

(b) QSML

10−2 10−1 100

0.1

0.2

0.3

0.4

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhoodqM > 0.1qM > 0.2Bifurc. & qM > 0.1Bifurc. & qM > 0.2

(c) SMC

10−2 10−1 100

0.1

0.2

0.3

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhood

(d) QSMC

Figure 9.7: ROC curves of the real-valued SML, SMC, QSML and QSMC type for the CASIA datasetwith applied minutiae pre-selection for determining a PSMR configuration.

10−3 10−2 10−1 100

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

SMC (Bifurc., Tuned)SMC TunedSMC

Figure 9.8: ROC curves of the most promising SMR types, tuned and untuned, for the CASIA dataset.Note: the tuned PSMC (SMC with only bifurcation minutiae pre-selection) and the untuned PSMCare the same ROC since the PSMC achieves the best biometric performance at λmax = 0.6.

very good contrast but strong noise4 and it is impossible to automatically compensate the ROItranslation. Through the latter, it is expected that the absolute-valued SMR performs betterthan the real-valued ones. Again, the input minutiae for the experiments are generated usingthe same feature extraction pipeline and settings as presented in table 9.1.

4Recall (section 8.1), there are hints of some additional, manual image enhancements that also boost thenoise.

85

Once again repeating the same experiments as for the PolyU and CASIA dataset above,the initial SML, SMC, QSML5 and QSMC plots are presented in figure 9.9. As expected, the

10−2 10−1 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

MHD-BaselineSMC (Absolute)SMC (Real)SML (Absolute)SML (Real)QSMC (Absolute)QSMC (Real)QSML (Absolute)QSML (Real)

Figure 9.9: ROC curves of the SML, SMC, QSML and QSMC type for the PUT dataset.

absolute-valued SMR yield better biometric performance than their real-valued counterparts.Again, the SMR performs better than the MHD, but the SMR still does not yield a accept-able performance. The performance gains achieved by applying the minutiae pre-selection ap-proaches to the PUT dataset, shown in figure 9.10, still do not yield desirable results. A deviationfrom the other datasets is the, in perspective, enormous performance gain of the PSMC (withapplied nighbourhood cleaning; figure 9.10c) compared to the SMC.

To complete the experiments for the PUT dataset, the last experiment is run to tune λ

for the SMR sampling as for the other datasets. Figure 9.11 shows the ROC of the tuned anduntuned SMR for the best five configurations yielded by the previous experiment (figure 9.10).

Finally, the top three configurations for the PUT dataset are summarized in table 9.6. Com-

SMR Ident. Verif.

Type Pre-Selection λmin λmax TP0.5 TP0 EER

PSML Bifurc. & qM > 0.2 0.1 0.68 41.8% 41.1% 8.3%PSML Bifurc. & qM > 0.1 0.1 0.47 41.4% 40.8% 8.5%SML - 0.1 0.68 43.3% 40.8% 8.5%

Table 9.6: Top three configurations, for a SMR biometric identification system using the PUT dataset,with their corresponding verification EER, ordered by TP0.

pared to the MHD, the SMR performs much better, although the results are still unacceptable.5Note: since the PUT features a much higher contrast than the other two datasets, the multiplier (gain)

for the quality information extracted with the maximum curvature algorithm was redetermined using the sameprocedure as the initial gain determination.

86

10010−2 10−1 100

0.1

0.2

0.3

0.4

0.5

0.6

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhoodqM > 0.1qM > 0.2Bifurc. & qM > 0.1Bifurc. & qM > 0.2

(a) SML

10−2 10−1 100

0.1

0.2

0.3

0.4

0.5

0.6

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhood

(b) QSML

10−2 10−1 100

0.1

0.2

0.3

0.4

0.5

0.6

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhoodqM > 0.1qM > 0.2Bifurc. & qM > 0.1Bifurc. & qM > 0.2

(c) SMC

10−2 10−1 100

0.1

0.2

0.3

0.4

0.5

0.6

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhood

(d) QSMC

Figure 9.10: ROC curves of the absolute-valued SML, SMC, QSML and QSMC type for the PUTdataset with applied minutiae pre-selection for determining a PSMR configuration.

10−3 10−2 10−1 100

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

SML (Bifurc. & qM > 0.2) TunedSML (Bifurc. & qM > 0.2)SML (Bifurc. & qM > 0.1) TunedSML (Bifurc. & qM > 0.1)SML (Bifurc.) TunedSML (Bifurc.)

Figure 9.11: ROC curves of the most promising SMR types, tuned and untuned, for the PUT dataset.

Again, these poor performances (especially MHD) hint that there might be a problem with thefeature extraction pipeline.

9.2.4 Aging / Further experiments

All presented datasets suffer from aging. Figure 9.12 exemplary shows the ROC for the PolyUdataset analysing intra-session comparison trials (both probe and reference stem from the samesession) versus inter-session comparison trials (probe and reference are not from the same

87

10−2 10−1 1000.6

0.7

0.8

0.9

1

False Positive Identification Rate

True

Posit

ive

Iden

tifica

tion

Rat

eQSML (Session 1)QSML (Session 2)QSML (Inter-Session)

Figure 9.12: ROC for images of different sessions from the PolyU dataset.

session). Refer to the discussion in chapter 10 for an explanation for these results.In order to continue the main experiments for this project on a reasonable basis, all followingexperiments will only be run with the first session of the PolyU dataset. However, all finalexperiments will be repeated twice for the PolyU dataset: once only using the second session,the second with an inter-session analysis.Refer to table A.1 for the top three configurations for the first session. The best result6 isyielded by the PSML with a minutiae pre-selection of qM > 0.2 (TP0 = 90.4%).

9.3 Workload reduction by feature reductionWith the best-performing SMR configuration, PSML with minutiae pre-selection by qM > 0.2(further just called PSML-Baseline), determined, the first workload reduction experiments areexecuted. This section begins by determining the workload baseline, followed by the PSML-CPCA, binary PSML and the binary PSML-CPCA.

9.3.1 Workload baseline for the PSML-pipeline

Recall (section 8.5), to state the workload reduction, a baseline workload is needed. The work-load is defined as ω = S ∗ p ∗ τ .For this experimental setup, 256 subjects are enrolled (S = 256)7. The naïve biometric systemcompares all enrolled templates (penetration rate (p) = 1.0). Recall (section 5.2), one PSML(and SMR in general) template comprises N×M = 32768 floating point numbers. There is nodeterministic way to calculate the cost (e.g. CPU cycle count) of one floating point operationon a modern x86 (x86_64) since it depends too much on e.g. accessing data in cache, branch

6In terms of TP0 and TP0.1, the highest TP0.5 = 92.0% is achieved by the QSML.7S is misleading in this context: since every user contributes two hands, 256 templates are enrolled from 128

subjects. In the PolyU it is not possible to link two biometric instances to one subject. Therefore it is assumed,that every subject presented one instance.

88

10−3 10−2 10−1 1000.8

0.85

0.9

0.95

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

PSML-BaselinePSML-CPCA L = 5PSML-CPCA L = 15PSML-CPCA L = 25PSML-CPCA L = 50PSML-CPCA L = 100

Figure 9.13: ROC for the CPCA feature reduction on the PolyU dataset with a training set of L = 5,15, 25, 50 and 100 PSML.

prediction and the instruction pipeline. Further, recall section 5.6.1: one SML comparison com-prises the sum of N×M floating point multiplications with one division, thus two operationsare actually executed per floating point: one multiplication and one addition.Therefore, it is assumed that these two floating point operations equal 32 bit-operations8. Thisassumption yields τ = 1048576.The final baseline workload is therefore ω = 256 ∗ 1 ∗ 1048576 ≈ 2.68×108.

9.3.2 CPCA

The first workload reduction approach uses the CPCA feature reduction. Applying the CPCAto the PSML-Baseline, yields little-to-no performance impairment, as shown in figure 9.13. Noperformance improvements could be achieved by using different SMR settings or CPCA train-ing. Refer to table A.2 for a full listing of TP0, TP0.1, TP0.5 and EER for different L, and totable 9.7. The CPCA applied to the PSML-Baseline is further called PSML-CPCA. Note: the

SMR Ident. Verif.

Label λmin λmax TP0.5 TP0.1 TP0 EER

PSML-Baseline 0.1 0.51 91.2% 91.1% 90.4% 2.1%PSML-CPCA L = 100 0.1 0.51 91.5% 90.6% 90.2% 2.1%

Table 9.7: TP0, TP0.1, TP0.5 and EER for the CPCA feature reduction on the PSML-Baseline.

experiment was run with L = 5, 15, 25, 50, 100 PSML for the CPCA. There are no mentionabledifferences between the training set sizes9. Therefore, even small databases with a very smalltraining set can apply the CPCA reduction for their SMR-based system and are able to train

8This follows the recommendation in [DRB17] to count the number of compared bits.9In fact, the training set with L = 15 performs better than all others in the means of TP0. The introduced

error depends more on the quality (e.g. the significance) of the templates used.

89

10−3 10−2 10−1 1000.8

0.85

0.9

0.95

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

PSML-BaselineBinary PSMLBinary PSML (Tuned: T P 0)Binary PSML (Tuned: T P 0.1)Binary PSML (Set-Bits/T P 0)

Figure 9.14: ROC for the binary PSML feature vector of the PSML-Baseline on the PolyU dataset.

a completely different CPCA, if necessary.

A additional minor SMR-CPCA configuration is the λCP CAmax , which states λmax for the SMR

that are used as CPCA training templates. Using a larger λCP CAmax then λmax results in a larger

SMR-CPCA projection necessary to keep the biometric performance at the level of the full-featured SMR. When using λCP CA

max = λmax with small λmax the SMR-CPCA templates mightbe reduced too much, which (drastically) lowers the biometric performance. If not stated oth-erwise, all following SMR-CPCA-based experiments are run with λCP CA

max = 0.6.

After the CPCA feature reduction with λCP CAmax = 0.6, it is safe to crop the PSML-CPCA

to MCP CA = 24 rows, yielding N×MCP CA = 6144, thus ω = 256 ∗ 1 ∗ (6144 ∗ 32) ≈ 5.03×107.This alone results in a workload reduction to 𝟋 ≈ 5.03×107

2.68×108 ≈ 0.1875 = 18.75%.

9.3.3 Binary SMR

A further workload reduction approach is the binarisation of the floating point feature vectorSMR. Reducing the feature vector from a 32bit to a 1bit feature vector results in a workloadof ω = 256 ∗ 1 ∗ 32768≈8.39×106, therefore achieving 𝟋 = 0.03125 = 3.125%. The binarisationexperiment was run in two stages: first, the performance has been evaluated using the PSML-Baseline settings with the recommended MT for the binarisation; and second, an exhaustiveexperiment has been run to determine the best performing settings for the binary PSML inthe means of λmax and mask thresholds. In the second experiment, different mask thresholdsfor the enrolled templates and the query templates have been used. As seen in figure 9.14 andstated in table 9.8, the binarisation step does not drastically impair the performance of theoriginal PSML-Baseline.

90

MT SMR Ident. Verif.

Label Enrol Query λmin λmax TP0.5 TP0.1 TP0 EER

Binary PSML 0.8 0.8 0.1 0.51 89.6% 87.9% 86.7% 2.4%Binary PSML (Tuned: TP 0) 0.7 0.6 0.1 0.47 89.6% 88.6% 88.2% 2.2%Binary PSML (Tuned: TP 0.1) 0.7 0.2 0.1 0.55 90.0% 89.4% 85.0% 2.3%Binary PSML (Set-Bits/TP 0) 0.6 0.6 0.1 0.51 89.4% 88.5% 87.7% 2.2%

Table 9.8: TP0 with corresponding TP0.1 and TP0.5 for a binary PSML biometric identification systemusing the PolyU dataset, with their corresponding verification EER, ordered by TP0.

10−3 10−2 10−1

0.8

0.85

0.9

0.95

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

PSML-BaselinePSML-CPCABinary PSML-CPCABinary PSML-CPCA (MaskT hreshold = 0.8)Binary PSML-CPCA (TP0 optimisation)Binary PSML-CPCA (MaskT hreshold = 0.9)Binary PSML-CPCA (MaskT hreshold = 0.3)

Figure 9.15: ROC for the binary PSML-CPCA feature vector on the PolyU dataset.

9.3.4 Binary SMR-CPCA

Combining the two feature reduction approaches above (CPCA and binarisation) results inan even greater workload reduction. Since MCP CA can be reduced to 20 for the binary PSML-CPCA, the feature vector of size N×MCP CA = 5120 comprises 1bit elements, ω = 256∗1∗5120 ≈1.31×106 is achieved, thus the workload reduction 𝟋 ≈ 0.004883 ≈ 0.4883%. Under consid-eration of the performance-lossless feature compression achieved by the CPCA and the smallamount of performance loss by the binarisation, the PSML-CPCA should achieve a comparableperformance to the binary PSML.The experiments confirm the previous assumption. Strictly speaking, the binary PSML-CPCA

performs better than the binary PSML, as seen in figure 9.15 and stated in table 9.9. Theexperiments has been run as in the previous experiment: first, the performance has been eval-uated using the best-performing PSML-CPCA settings with the recommended mask thresholdfor the binarisation (MT = 0.8); and second, an exhaustive experiment has been run to de-termine the best performing settings for the binary PSML-CPCA in the means of λmax andmask thresholds. Additionally reported results feature a TP0 optimised configuration, a lowmask threshold configuration (MT = 0.3) and a high threshold, thus theoretically most stableon the edges, configuration (MT = 0.9).

91

MT SMR Ident. Verif.

Label Enrol Query λmin λmax TP0.5 TP0.1 TP0 EER

Binary PSML-CPCA TP0 0.6 0.4 0.1 0.36 89.7% 89.1% 89.0% 2.2%Binary PSML-CPCA MT = 0.9 0.9 0.9 0.1 0.45 90.0% 89.0% 88.4% 2.2%Binary PSML-CPCA MT = 0.3 0.3 0.3 0.1 0.51 88.9% 88.3% 86.9% 2.3%Binary PSML-CPCA MT = 0.8 0.8 0.8 0.1 0.51 90.0% 88.4% 86.7% 2.3%

Table 9.9: TP0 with corresponding TP0.1 and TP0.5 for a binary PSML-CPCA biometric identificationsystem using the PolyU dataset, with their corresponding verification EER, ordered by TP0.

9.3.5 Summary

With the feature reduction approaches, a significant workload reduction is already achieved.The smallest workload reduction 𝟋≈0.4883% is reached by the binary PSML-CPCA with a re-lative TP0 recognition performance loss of only 1.4%-points compared to the PSML-Baseline.Looking at the workload reduction achieved, the losses of the binary approaches are accept-

Workload Ident. Verif.

Approach ω256 𝟋 TP0.5 TP0.1 TP0 EERTP0 loss to

PSML-Baseline

PSML-Baseline 2.68×108 − 91.2% 91.1% 90.4% 2.1% −PSML-CPCA 5.03×107 18.748% 91.5% 90.6% 90.2% 2.1% 0.2%Binary PSML 8.39×106 3.125% 89.6% 88.6% 88.2% 2.2% 2.2%Binary PSML-CPCA 1.31×106 0.488% 89.7% 89.1% 89.0% 2.2% 1.4%

Table 9.10: Summary of achieved TP0 with corresponding TP0.1 and TP0.5 for SMR biometric identi-fication system using the PolyU dataset and their corresponding verification EER with applied featurereduction approaches, ordered by TP0.

able. Furthermore, both Bloom filter-indexing and CPCA-Tree-indexing require the binaryrepresentation. Therefore, the next experiments base on the results of this section. The nextsections will explore the capabilities of two indexing approaches to reach a workload reductionby penetration rate reduction.

9.4 Bloom filter-indexing ApproachThe Bloom filter-indexing approach applied in a iris recognition application achieved a workloadreduction down to 𝟋≈1% in [DRB17]. However, the binary SMR shows different propertiescompared to the iris-code. This section presents the results achieved with the Bloom filter forthe binary PSML-CPCA, starting with the main identification experiments, followed by a briefoverview of the Bloom filter’s verification capabilities.

92

9.4.1 Binary SMR-CPCA Bloom filter-indexing

The binary PSML-CPCA achieved better results than the binary PSML, in terms of both re-garding biometric performance and workload reduction. Therefore, the binary PSML-CPCA ischosen as the first input for the Bloom filter-indexing.

First, the best-performing H and W need to be found. Since the PSML-CPCA is compar-atively small, Bloom filter block sizes 4 ≤ H ≤ 8 respectively 4 ≤ W ≤ 16 are tested. Forevery H and W pair, the binarisation mask thresholds 0.1, 0.2, . . . , 0.9 are used to find theoptimal mask-block size combination. To obtain the most clean results for finding H and W inidentification scenarios, a Bloom filter tree configuration with no workload reduction is chosen;T = 64, without quick traversal decision and without tree pre-selection.The results of the first Bloom filter experiments are listed in table 9.11. Even with no workload

Bloom filter SMR Workload Recognition Performance

W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

4 5 0.1 2.09 ∗ 106 0.7812% 160.0% 82.1% 80.9% 79.6% 10.8%4 5 0.2 2.09 ∗ 106 0.7812% 160.0% 82.1% 78.0% 76.6% 13.8%5 4 0.5 1.04 ∗ 106 0.3906% 80.0% 80.2% 78.0% 76.6% 13.8%4 5 0.3 2.09 ∗ 106 0.7812% 160.0% 79.1% 77.7% 76.6% 13.8%4 4 0.4 1.31 ∗ 106 0.4882% 100.0% 80.9% 76.8% 73.6% 16.8%5 4 0.6 1.04 ∗ 106 0.3906% 80.0% 77.1% 75.4% 73.5% 16.9%6 4 0.7 8.73 ∗ 105 0.3256% 66.7% 76.9% 75.9% 72.3% 18.1%6 4 0.8 8.73 ∗ 105 0.3256% 66.7% 74.5% 71.8% 68.5% 21.9%7 4 0.9 7.48 ∗ 105 0.2790% 57.1% 70.3% 69.2% 69.2% 21.2%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table 9.11: Best achieved binary PSML-CPCA Bloom filter TP0 with corresponding TP0.1 and TP0.5

for every tested binarisation mask threshold with their corresponding H and W (T = 64), ordered byTP0,TP0.1.

reduction by using lesser trees, the Bloom filter loses ∼10.8%-points of TP0 in its best con-figuration for the binary PSML-CPCA. Furthermore, the best configuration reintroduces 60%workload compared to the naïve binary PSML-CPCA approach, due to the small Bloom filterblock size. Using a larger Bloom filter block size drastically reduces the achieved performance.

The origin of the poor performance of the Bloom filter applied to binary PSML-CPCAtemplates is the high number of bit errors when comparing two mated binary PSML-CPCA,visualised in figure 9.16. Even at a very high mask threshold of 0.9, the bit-error rate is 12.4%with an error in more than 50% of columns, which is excessive for a reliable Bloom filter trans-

93

(a) Threshold 0.1 (b) Threshold 0.4 (c) Threshold 0.9

Figure 9.16: Bit-errors between two mated binary PSML-CPCA templates at three different binarisa-tion mask thresholds (white = bit-error).

formation that needs stable column-bits. Several bit-error correction methods for more robustcolumns were tried to achieve a higher recognition performance for the PSML-CPCA Bloomfilter: majority-voting (3 rows), majority-voting (5 rows), majority-voting (N4), majority-voting(N8) and a median filter. Every correction method was also tested with a transposed PSML-CPCA. Among all correction methods, the majority-voting over 3 rows achieved the highestperformance, albeit much less than without bit-error correction. Appended table A.3 lists theresults of the majority-voting, the transposed PSML-CPCA and the transposed PSML-CPCAwith majority-voting.

While analysing the poor performance of the Bloom filter for the binary PSML-CPCA, itwas noticeable that the pre-selection error rate (PER; rate of pre-selection error events) wassmall, but their genuine comparison scores overlapped too much with the scores of impostorqueries. Therefore, another experiment was run to test whether an additional comparisonbetween the original enrolled binary PSML-CPCA template of the retrieved candidate and theoriginal query binary PSML-CPCA template could increase the biometric performance whilesacrificing10 a minimal amount of workload reduction.As listed in table 9.12, employing an additional PSML-CPCA comparison increases the per-

formance up to 15% points. In some configurations, the generated overhead of the additionalcomparison is compensated by the increased Bloom filter block size (used to achieve a lowerPER) compared to the previous experiment. Unfortunately, in the best-performing configura-tion the workload is doubled compared to the naïve PSML-CPCA.

However, the Bloom filter in combination with the binary search tree approach is able toreduce the workload by reducing the number of trees built and using the tree pre-selection.The next experiment is a repetition of the previous experiment with halved and quartered treecount (T = 32, T = 16, no tree pre-selection, additional PSML-CPCA comparison). As listedin table 9.13, reducing T to 1/16 of enrolled templates introduces a very high performance losswhile hardly reducing the workload compared to 1/8 of enrolled templates due to too many bitcollisions and overly-populated tree nodes for a correct traversal direction decision. Note: oneconfiguration T = 32 (W = 5,H = 4, mask threshold = 0.7) yields a higher biometric perform-ance than all T = 64 results while achieving a much higher workload reduction. Comparing all

10Recall, the number of bit comparisons for a Bloom filter template equals its size: 2H ∗MCP CA/H ∗N/W; thusω256 = 256 ∗ 2H ∗MCP CA/H ∗N/W. With the additional binary PSML-CPCA comparison, the workload increasesto ω256 = 256 ∗ 2H ∗MCP CA/H ∗ N/W + MCP CA ∗N .

94

Bloom filter SMR Workload Recognition Performance

W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

5 6 0.4 2.80 ∗ 106 1.044% 213.7% 87.3% 87.1% 86.2% 4.2%6 5 0.9 1.40 ∗ 106 0.523% 107.1% 86.9% 86.2% 85.8% 4.6%6 6 0.6 2.33 ∗ 106 0.870% 178.2% 87.3% 86.6% 85.6% 4.8%4 7 0.2 5.99 ∗ 106 2.234% 457.5% 86.6% 85.5% 85.2% 5.2%6 6 0.3 2.33 ∗ 106 0.870% 178.2% 87.1% 86.1% 84.9% 5.5%5 8 0.5 8.39 ∗ 106 3.127% 640.4% 83.9% 83.6% 83.0% 7.4%7 7 0.7 3.42 ∗ 106 1.277% 261.6% 83.7% 83.6% 83.2% 7.2%5 6 0.8 2.80 ∗ 106 1.044% 213.7% 85.8% 85.3% 83.7% 6.7%4 7 0.1 5.99 ∗ 106 2.234% 457.5% 85.9% 84.9% 83.7% 6.7%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table 9.12: Best achieved binary PSML-CPCA Bloom filter TP0 with corresponding TP0.1 and TP0.5,with a final binary PSML-CPCA comparison, for every tested binarisation mask threshold with theircorresponding H and W (T = 64), ordered by TP0,TP0.1.

Bloom filter SMR Workload Recognition Performance

T W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

16

4 4 0.2 6.60 ∗ 105 0.246% 50.4% 84.5% 83.8% 83.0% 7.4%4 6 0.4 1.75 ∗ 106 0.653% 133.7% 83.0% 82.7% 82.5% 7.90%4 6 0.5 1.75 ∗ 106 0.653% 133.7% 83.0% 82.5% 82.3% 8.1%

. . . ‡

32

5 4 0.7 7.91 ∗ 105 0.295% 60.4% 87.8% 87.6% 86.6% 3.8%5 5 0.8 1.26 ∗ 106 0.471% 96.4% 87.7% 87.3% 86.2% 4.2%8 5 0.4 7.91 ∗ 105 0.295% 60.4% 87.0% 86.6% 86.1% 4.3%

. . . ‡

* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.4.

Table 9.13: Top three achieved binary PSML-CPCA Bloom filter TP0 with corresponding TP0.1 andTP0.5, with a final binary PSML-CPCA comparison using T = 32 and T = 16 trees.

other T = 32 and T = 64 configurations, this could be serendipity.

Since the biometric performance loss introduced by using T less than 1/8 of enrolled tem-plates is too high, another T = 64 and T = 32 experiment has been run with enabled treepre-selection. Tree pre-selection introduces additional workload by first comparing all tree rootnodes with the query template, thus selecting more than 48 trees for T = 64 and more than

95

26 trees for T = 32 would actually increase the workload. Observe table 9.14, reducing the

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 4 4 0.2 3.53 ∗ 105 0.1316% 26.9% 78.2% 77.3% 77.1% 13.3%2 4 4 0.3 3.73 ∗ 105 0.1392% 28.5% 82.0% 81.4% 81.3% 9.1%3 4 4 0.4 3.94 ∗ 105 0.1469% 30.1% 84.7% 84.3% 83.4% 7.0%

. . . ‡

18 5 5 0.6 8.96 ∗ 105 0.3339% 68.4% 89.2% 88.6% 87.7% 2.7%19 5 5 0.6 9.22 ∗ 105 0.3437% 70.4% 89.5% 88.9% 88.0% 2.4%20 5 5 0.6 9.48 ∗ 105 0.3535% 72.4% 89.5% 88.9% 88.0% 2.4%

. . . ‡

46 5 6 0.4 2.71 ∗ 106 1.0113% 207.1% 87.3% 87.1% 86.2% 4.2%47 5 6 0.4 2.75 ∗ 106 1.0278% 210.4% 87.3% 87.1% 86.2% 4.2%48 5 6 0.4 2.80 ∗ 106 1.0444% 213.7% 87.3% 87.1% 86.2% 4.2%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.5.

Table 9.14: Best achieved binary PSML-CPCA Bloom filter TP0 with corresponding TP0.1 and TP0.5,with a final binary PSML-CPCA comparison and tree pre-selection of t = 1, . . . , 48 trees at T = 64.

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 4 5 0.1 3.16 ∗ 105 0.1179% 24.1% 66.6% 65.9% 65.7% 24.7%2 4 5 0.1 3.65 ∗ 105 0.1362% 27.9% 74.2% 73.8% 73.2% 17.2%3 4 5 0.3 4.14 ∗ 105 0.1545% 31.6% 76.6% 76.2% 76.0% 14.4%

. . . ‡

16 4 4 0.5 6.60 ∗ 105 0.2460% 50.4% 87.7% 86.7% 86.6% 3.8%17 4 5 0.6 1.10 ∗ 106 0.4108% 84.1% 88.1% 87.2% 86.7% 3.7%18 4 5 0.6 1.15 ∗ 106 0.4292% 87.9% 88.1% 87.2% 86.7% 3.7%

. . . ‡

26 5 4 0.7 7.75 ∗ 105 0.2888% 59.1% 87.4% 87.2% 86.2% 4.2%27 5 4 0.7 7.99 ∗ 105 0.2979% 61.0% 87.5% 87.3% 86.3% 4.1%28 5 4 0.7 8.24 ∗ 105 0.3071% 62.9% 87.6% 87.3% 86.4% 4.0%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.6.

Table 9.15: Best achieved binary PSML-CPCA Bloom filter TP0 with corresponding TP0.1 and TP0.5,with a final binary PSML-CPCA comparison and tree pre-selection of t = 1, . . . , 26 trees at T = 32.

96

number of pre-selected trees to 19 to 21 increased the performance while reducing the workload∼30%. In its best performance, the Bloom filter with T = 32 suffers a ∼1.3%-points perform-ance loss compared to the T = 64 setup. Because a smaller Bloom filter block size is needed toreceive this performance for T = 32 with tree pre-selection (table 9.15), the workload reductionis less11 than the workload reduction with T = 64 and tree pre-selection.

The binary PSML-CPCA Bloom filter achieved the best biometric performance of TP0 =88% with T = 64, a tree pre-selection of t≈5/16 and an additional binary PSML-CPCA com-parison. This is a TP0 loss of 1%-point compared to the binary PSML-CPCA baseline whilereaching a workload reduction of just 30%. One final experiment is run to test whether usingthe real-valued PSML-CPCA as the additional comparison is able to achieve a higher biomet-ric performance compared to the binary PSML-CPCA baseline while receiving at least a smallamount of workload reduction, even if the comparison cost for a real-valued PSML-CPCA is32 times higher than the binary PSML-CPCA comparison. Since the three template protectionrequirements are proven in section 7.2, it is safe to keep these templates. As listed in table 9.16,

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 4 4 0.2 3.53 ∗ 105 0.1316% 26.95% 78.9% 78.4% 78.0% 12.4%2 4 5 0.3 5.94 ∗ 105 0.2216% 45.39% 83.6% 83.1% 82.7% 7.7%3 4 4 0.4 3.94 ∗ 105 0.1469% 30.08% 85.4% 84.7% 84.6% 5.8%

. . . ‡

25 6 4 0.5 5.64 ∗ 105 0.2104% 43.1% 89.8% 89.1% 89.0% 1.4%26 6 4 0.5 5.78 ∗ 105 0.2155% 44.14% 89.9% 89.1% 89.0% 1.4%27 6 4 0.5 5.92 ∗ 105 0.2206% 45.18% 90.0% 89.1% 89.0% 1.4%

. . . ‡

46 6 5 0.6 1.35 ∗ 106 0.5065% 103.7% 90.2% 89.4% 88.8% 1.6%47 4 5 0.6 2.06 ∗ 106 0.7710% 157.9% 90.0% 89.4% 88.8% 1.6%48 4 5 0.6 2.10 ∗ 106 0.7832% 160.4% 90.0% 89.4% 88.8% 1.6%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.7.

Table 9.16: Best achieved binary PSML-CPCA Bloom filter TP0 with corresponding TP0.1 and TP0.5,with a final real-valued PSML-CPCA comparison and tree pre-selection of t = 1, . . . , 48 trees atT = 64.

the additional real-valued PSML-CPCA comparison achieves a higher biometric performance,than the binary PSML-CPCA baseline.

11When sacrificing a another small amount of biometric performance, a workload reduction ∼50% is achievablewith T = 32.

97

This and the two previous experiments use the Bloom filter-indexing as a rank 1 pre-selectionapproach with one final real-valued PSML-CPCA, respective one final binary PSML-CPCAcomparison. With a tree pre-selection of t = 25 trees at T = 64 and with the comparisoncost of the real-valued comparison calculated, an additional12 workload reduction of ∼57% isreached at TP0 = 89% - just ∼1.42%-points less than the heavy, naïve PSML approach, whileachieving the same biometric performance as the naïve binary PSML-CPCA approach andoutperforming all previous binary PSML-CPCA Bloom filter-indexing approaches in terms ofboth biometric performance and workload.

9.4.2 Binary SMR Bloom filter-indexing

Bloom filter-indexing with binary PSML-CPCA templates yielded in their best performing con-figuration a TP0 of 88% with an additional workload reduction of ∼30%. Workload reductionsof ∼50% are achieved with a TP0 of 86%. The following experiments are run to test whetherthe binary PSML is able to achieve a higher biometric performance than the smaller binaryPSML-CPCA, although the base performance is lower.

Again, the best-performing H andW need to be found. The experiment is run the same wayas for the binary PSML-CPCA Bloom filter. In the most basic experiment, the binary PSML

Bloom filter SMR Workload Recognition Performance

W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

5 4 0.2 6.71 ∗ 106 2.500% 512.0% 85.3% 83.6% 82.5% 7.9%5 4 0.1 6.71 ∗ 106 2.500% 512.0% 85.5% 82.7% 81.2% 9.2%6 4 0.3 5.59 ∗ 106 2.083% 426.9% 83.9% 82.7% 80.5% 9.9%7 4 0.4 4.79 ∗ 106 1.786% 364.8% 82.4% 80.5% 79.9% 10.5%8 6 0.5 1.11 ∗ 107 4.167% 853.2% 82.0% 80.5% 79.0% 11.4%7 4 0.6 4.79 ∗ 106 1.786% 365.4% 80.9% 77.0% 76.3% 14.1%7 4 0.7 4.79 ∗ 106 1.786% 365.4% 78.7% 76.1% 72.1% 18.3%8 8 0.8 3.35 ∗ 107 12.500% 2560.0% 71.2% 70.2% 66.2% 24.2%8 8 0.9 3.35 ∗ 107 12.500% 2560.0% 67.3% 61.9% 52.7% 37.7%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table 9.17: Best achieved binary PSML Bloom filter TP0 with corresponding TP0.1 and TP0.5 forevery tested binarisation mask threshold with their corresponding H and W (T = 64), ordered byTP0,TP0.1.

is able to achieve higher TP0 values (table 9.17) than the binary PSML-CPCA, although the12Workload reduction on top of the workload-reduction-by-feature-reduction.

98

performance loss compared to the baseline is too high.

By inspecting the bit-errors between two mated binary PSML templates, it is clearly visiblethat the binary PSML also suffers from a high bit-error rate. Even at a high mask thresholdof 0.9, a bit-error rate of ∼14% is recorded. Figure 9.17 exemplary shows the bit-error betweenmated templates for different mask thresholds.

(a) Threshold 0.1 (b) Threshold 0.4 (c) Threshold 0.9

Figure 9.17: Bit-errors between two mated binary PSML templates from one subject (i.e. one instance)at three different binarisation mask thresholds (white = bit-error).

Therefore, the same bit-error correction methods as in the binary PSML-CPCA experi-ments are tested for the binary PSML. Both transposing the PSML and using majority voting

Bit-error Bloom filter SMR Workload Recognition Performance

correction W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

Transposed

4 5 0.1 1.34 ∗ 107 5.000% 1024.0% 84.5% 84.1% 83.0% 7.4%4 4 0.2 8.38 ∗ 106 3.125% 640.0% 84.4% 82.8% 82.1% 8.3%6 4 0.3 5.59 ∗ 106 2.083% 426.9% 83.8% 81.9% 81.6% 8.8%

. . . ‡

Majority Voting

5 4 0.1 6.71 ∗ 106 2.500% 512.0% 84.7% 83.0% 81.6% 8.8%5 4 0.2 6.71 ∗ 106 2.500% 512.0% 84.7% 84.0% 81.6% 8.8%6 4 0.3 5.59 ∗ 106 2.083% 426.9% 84.1% 82.0% 80.9% 9.5%

. . . ‡

Transposed&

Majority Voting

4 5 0.1 1.34 ∗ 107 5.000% 1024.0% 84.3% 83.4% 82.9% 7.5%4 4 0.2 8.38 ∗ 106 3.125% 640.0% 84.8% 82.9% 82.3% 8.1%6 6 0.3 1.49 ∗ 107 5.556% 1137.9% 82.3% 81.3% 79.7% 10.7%

. . . ‡

* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.8.

Table 9.18: Top three achieved TP0 with corresponding TP0.1 and TP0.5 for every tested binarisationmask threshold with their corresponding H and W (T = 64) using the binary PSML with severalbit-error reduction strategies, ordered by mask threshold.

increased the biometric performance, but transposing and using majority voting fails to increasethe biometric performance compared to only using the transposed PSML. Unfortunately, trans-posing the binary PSML introduces additional workload due to the Bloom filter transformation

99

process.

Since the error introduced by the Bloom filter is still too high with the bit-error correctionmethods, the additional comparison strategy as for the binary PSML-CPCA is tested for thenormal, transposed and majority voting binary PSML. As listed in table 9.19, with enabled

Bit-error Bloom filter SMR Workload Recognition Performance

correction W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

None

8 6 0.6 1.12 ∗ 107 4.179% 855.6% 89.8% 89.5% 89.0% 1.4%7 7 0.7 2.19 ∗ 107 8.175% 1674.2% 90.1% 89.1% 88.9% 1.5%9 4 0.5 3.76 ∗ 106 1.401% 286.7% 89.2% 88.8% 88.8% 1.6%

. . . ‡

Transposed

6 6 0.5 1.49 ∗ 107 5.568% 1140.4% 89.4% 88.9% 88.5% 1.9%5 5 0.6 1.07 ∗ 107 4.012% 821.7% 89.8% 88.7% 88.0% 2.3%5 8 0.7 5.37 ∗ 107 20.008% 4098.4% 89.8% 88.8% 87.9% 2.5%

. . . ‡

MajorityVoting

8 5 0.8 6.74 ∗ 106 2.512% 514.5% 89.8% 89.5% 88.9% 1.5%7 4 0.7 4.82 ∗ 106 1.798% 368.0% 90.2% 89.4% 88.6% 1.8%6 5 0.6 8.98 ∗ 106 3.346% 685.4% 90.1% 89.1% 88.5% 1.9%

. . . ‡

* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.9.

Table 9.19: Top three achieved TP0 with corresponding TP0.1 and TP0.5 using the binary PSML withand without bit-error reduction strategies, ordered by TP0.

additional comparison and without bit-error reduction strategies, a biometric performance ofup to TP0 = 89% is reached (0.9%-points more than the binary PSML baseline, equals the bin-ary PSML-CPCA baseline) but with an increase of ∼33% in workload. Sacrificing 0.2% pointsin TP0, a workload reduction of ∼55% is reached. With 𝟋 = 0.701%, the overall workloadreduction is still smaller than the workload reduction of the binary PSML-CPCA Bloom filter(𝟋 = 0.172%, TP0 = 88%).

All previous binary SMR Bloom filter experiments have been run at T = 64. The experi-ments were also run with T = 32 and T = 16, with their results listed in table 9.20. Again,using T = 16 introduces too much error in the means of TP0. Unfortunately, the best con-figuration with T = 32 fails to achieve a much higher workload reduction then the third bestT = 64 configuration and loses ∼0.6%-points TP0.

Finally, two last experiments are run at T = 32 and T = 64 with enabled tree pre-selection.Using the final binary PSML comparison with tree pre-selection at T = 64 achieves a biometricperformance of TP0 = 88.9% with a pre-selection of t = 5/16 trees (table 9.21). Compared tothe naïve binary PSML approach, this is an TP0 increase of 0.9% and an additional workload

100

Bloom filter SMR Workload Recognition Performance

T W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

16

5 5 0.5 5.40 ∗ 106 2.012% 412.1% 82.4% 81.8% 81.6% 8.8%4 5 0.6 6.74 ∗ 106 2.512% 514.5% 84.1% 83.8% 83.3% 7.1%4 6 0.7 1.12 ∗ 107 4.179% 855.6% 83.3% 82.5% 81.9% 8.5%

. . . ‡

32

7 4 0.7 3.62 ∗ 106 1.351% 277.1% 89.0% 88.1% 88.0% 2.3%7 4 0.5 3.62 ∗ 106 1.351% 277.1% 89.2% 88.4% 87.9% 2.5%5 4 0.6 5.06 ∗ 106 1.887% 386.5% 89.1% 88.5% 87.8% 2.6%

. . . ‡

* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.10.

Table 9.20: Top three achieved binary PSML Bloom filter TP0 with corresponding TP0.1 and TP0.5,with a final binary PSML comparison using T = 32 and T = 16 trees.

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 4 5 0.1 3.59 ∗ 106 1.340% 274.5% 72.1% 72.0% 71.7% 18.7%2 4 5 0.2 3.80 ∗ 106 1.418% 290.5% 79.1% 78.7% 78.0% 12.4%3 4 4 0.6 2.52 ∗ 106 0.940% 192.6% 82.9% 82.1% 81.2% 9.2%

. . . ‡

19 7 4 0.7 2.65 ∗ 106 0.989% 202.2% 89.6% 88.8% 88.8% 1.6%20 7 4 0.7 2.72 ∗ 106 1.017% 208.0% 89.8% 89.0% 88.9% 1.5%21 4 7 0.6 2.22 ∗ 107 8.271% 1694.0% 89.8% 89.4% 88.6% 1.8%

. . . ‡

35 7 7 0.7 1.74 ∗ 107 6.517% 1335.0% 90.1% 89.1% 88.9% 1.5%. . . ‡

44 8 6 0.6 1.05 ∗ 107 3.918% 802.5% 89.8% 89.5% 89.0% 1.4%. . . ‡

* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.11.

Table 9.21: Best achieved binary PSML Bloom filter TP0 with corresponding TP0.1 and TP0.5, witha final binary PSML comparison and tree pre-selection of t = 1, . . . , 48 trees at T = 64.

reduction down to ∼70%. The configuration with T = 32 (table 9.22) fails to yield equal TP0

values at a comparable workload reduction.

However, the binary PSML Bloom filter does not accomplish the overall workload reduc-

101

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 4 5 0.4 2.02 ∗ 106 0.754% 154.2% 52.3% 52.2% 51.9% 38.5%2 5 5 0.8 1.87 ∗ 106 0.699% 143.4% 62.7% 62.5% 62.5% 27.9%3 5 5 0.8 2.12 ∗ 106 0.793% 162.6% 70.0% 69.8% 69.8% 20.6%

. . . ‡

15 4 5 0.6 6.42 ∗ 106 2.395% 490.2% 88.4% 87.7% 86.7% 3.7%16 4 5 0.6 6.74 ∗ 106 2.512% 514.5% 88.6% 88.0% 87.0% 3.4%17 4 5 0.6 7.05 ∗ 106 2.629% 538.2% 88.7% 88.0% 87.0% 3.4%

. . . ‡

24 7 4 0.7 3.32 ∗ 106 1.240% 254.1% 88.8% 87.9% 87.8% 2.6%25 7 4 0.7 3.44 ∗ 106 1.282% 262.4% 88.8% 87.9% 87.8% 2.6%26 7 4 0.7 3.55 ∗ 106 1.324% 270.7% 88.8% 88.0% 87.9% 2.5%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.12.

Table 9.22: Best achieved binary PSML Bloom filter TP0 with corresponding TP0.1 and TP0.5, witha final binary PSML comparison and tree pre-selection of t = 1, . . . , 26 trees at T = 32.

tion13 and biometric performance of the binary PSML-CPCA Bloom filter. Using an additionalreal-valued PSML comparison as in the binary PSML-CPCA Bloom filter experiments is notfeasible: the real-valued PSML does not fulfil the unlinkability and renewability template pro-tection requirements. Therefore, keeping real-valued PSML is a privacy issue, and thus is irre-sponsible for real-world scenarios.For the sake of completeness, the appended table A.13 lists the results of such approach. Witha biometric performance of TP0 = 90.2%14 at a relative workload reduction of ∼64% comparedto the naïve binary PSML approach, this approach outperforms every other binary and Bloomfilter approach in the means of biometric performance, but the workload is still approximately5 times higher than the best binary PSML-CPCA Bloom filter and approximate 2 times higherthan the naïve binary PSML-CPCA approach. Even an increase in TP0 of 0.1% compared to thereal-valued PSML-Baseline is achieved with 𝟋B = 6.653%. At least this result hints at a con-firmation of the at-least-one-false-match problem from section 2.2.3: the biometric performanceincreased by reducing the workload.

9.4.3 Binary SMR-CPCA Bloom filter-verification

In [DRB17], the Bloom filter achieved the best verification performance with comparativelysmall block sizes. Looking at the biometric performance for the indexing approach with a small

13Workload reduction compared to the real-valued PSML-Baseline.14When doubling the workload due to a bad Bloom filter block size, a TP0 of 90.4% is reached.

102

Bloom filter block size, it has to be expected that the Bloom filter will not be feasible forverification purposes with binary PSML-CPCA templates.As shown in table 9.23, the best block size for verification purposes isW = 4,H = 2, resulting

For Bloom filter SMR Workload Recognition Performance

every W H MT ω1 𝟋B* 𝟋R

† EER EER Loss to PSLM-Baseline

H

4 2 0.5 2.56 ∗ 103 0.244% 50.00% 2.8% 0.7%4 3 0.5 3.41 ∗ 103 0.326% 66.7% 3.2% 1.1%4 4 0.2 5.12 ∗ 103 0.488% 100.0% 3.3% 1.2%5 5 0.2 6.55 ∗ 103 0.625% 128.0% 3.2% 1.1%5 6 0.3 1.09 ∗ 104 1.042% 213.3% 4.2% 2.1%4 7 0.2 2.34 ∗ 104 2.232% 457.1% 4.9% 2.8%4 8 0.5 4.09 ∗ 104 3.906% 800.0% 5.2% 3.1%

W

2 2 0.5 5.12 ∗ 103 0.488% 100.0% 3.0% 0.9%3 2 0.5 3.41 ∗ 103 0.326% 66.7% 3.0% 0.9%4 2 0.5 2.56 ∗ 103 0.244% 50.0% 2.8% 0.7%5 2 0.5 2.04 ∗ 103 0.195% 40.0% 3.0% 0.9%6 2 0.6 1.70 ∗ 103 0.163% 33.3% 3.1% 1.0%7 2 0.6 1.46 ∗ 103 0.140% 28.6% 3.1% 1.0%8 2 0.6 1.28 ∗ 103 0.122% 25.0% 3.2% 1.1%

* 𝟋 compared to the PSML ω1.† 𝟋 compared to the binary PSML-CPCA ω1.

Table 9.23: Best achieved EER by applying the Bloom filter to the binary PSML-CPCA, sorted byeach W and H.

in a workload reduction of 50% with an EER loss of 0.9%-points compared to the real-valuedPSML-Baseline and real-valued PSML-CPCA. Looking upon the score distribution, for the in

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.1

0.2

0.3

Similarity Score

Den

sity

Mated Scores

Non-mated Scores

(a) Bloom filter (W = 4, H = 2)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.1

0.2

0.3

Similarity Score

Den

sity

Mated Scores

Non-mated Scores

(b) binary PSML-CPCA

Figure 9.18: Verification score distribution (Y-axis normalised to 1) of the Bloom filter configurationachieving the best EER (W = 4,H = 2) for the binary PSML-CPCA in comparison to the scoredistribution of the binary PSML-CPCA.

103

table 9.23 highlighted configuration, in figure 9.18, the bad EER is easily explainable by thestrong overlap of the genuine and impostor scores.

9.4.4 Binary SMR Bloom filter-verification

The binary PSML achieved a higher biometric identification performance in the basic Bloomfilter approach. Therefore, it is expected that the binary PSML also results in a higher bio-metric verification performance. The results of the verification experiment, listed in table 9.24,

For Bloom filter SMR Workload Recognition Performance

every W H MT ω1 𝟋B* 𝟋R

† EER EER Loss to PSLM-Baseline

H

3 2 0.1 2.18 ∗ 104 2.083% 426.9% 2.50% 0.40%2 3 0.1 4.37 ∗ 104 4.167% 853.1% 2.68% 0.60%3 4 0.1 4.37 ∗ 104 4.167% 853.1% 2.73% 0.63%2 5 0.1 1.05 ∗ 105 10.000% 2047.9% 2.76% 0.66%5 6 0.1 6.99 ∗ 104 6.667% 1365.1% 2.94% 0.84%2 7 0.1 3.00 ∗ 105 28.567% 5851.3% 2.89% 0.79%2 8 0.1 5.24 ∗ 105 50.000% 10239.6% 3.13% 1.03%

W

2 2 0.1 3.28 ∗ 104 3.125% 640.0% 2.53% 0.43%3 2 0.1 2.18 ∗ 104 2.083% 426.9% 2.50% 0.40%4 2 0.1 1.64 ∗ 104 1.562% 320.0% 2.61% 0.51%5 2 0.1 1.31 ∗ 104 1.250% 256.0% 2.68% 0.58%6 2 0.1 1.09 ∗ 104 1.042% 213.1% 2.80% 0.70%7 2 0.2 9.36 ∗ 103 0.893% 183.0% 2.89% 0.79%8 3 0.3 1.09 ∗ 104 1.042% 213.1% 3.03% 0.93%

* 𝟋 compared to the PSML ω1.† 𝟋 compared to the binary PSML-CPCA ω1.

Table 9.24: Best achieved EER by applying the Bloom filter to the binary PSML, sorted by each Wand H.

confirm these expectations. Looking at the score distribution in figure 9.19 for the highlightedconfiguration, the genuine and impostor comparison scores are more separated than those forthe binary PSML-CPCA Bloom filter, thus achieving the lower EER, but still featuring a highoverlap. However, in consideration of the achieved workload reduction15 and the EER, applyingthe Bloom filter on the binary SMR for verification purposes is not feasible at least for highnoise environments.

9.5 CPCA-Tree-indexing ApproachApplying the Bloom filter to the binary PSML-CPCA with additional real-valued comparisonoutperformed the naïve binary PSML-CPCA approach in terms of both workload and biometric

15The workload ω does not consider the cost of transforming the binary input in the new Bloom filterrepresentation, therefore the real workload is a fraction higher.

104

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.1

0.2

0.3

Score

Den

sity

Mated Scores

Non-mated Scores

Figure 9.19: Verification score distribution of the Bloom filter configuration achieving the best EER(W = 4,H = 2) for the binary PSML.

performance. The next experiments are run to determine whether the introduced CPCA-Treeis able to outperform the Bloom filter approach in terms of workload reduction or biometricperformance in one of its three types.

9.5.1 Basic implementation

In the basic configuration, the CPCA-Tree is built using the binary PSML-CPCA created withthe high mask threshold configuration from section 9.3.4. As already discussed in section 7.1,employing more than 8 templates per tree introduces a very high population count in the (root)nodes, and thus is not feasible. Therefore, the experiments are run with T = 32 and T = 64.

Construction SMR Workload Recognition Performance

Type T λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

PSML-CPCA-Applied (PSML-CPCA-A)

640.49 0.6 0.65 1.31 ∗ 106 0.4883% 100.0% 4.141% 86.5% 85.4% 84.1% 6.3%0.48 0.6 0.65 1.31 ∗ 106 0.4883% 100.0% 4.062% 84.7% 83.4% 83.4% 7.0%0.47 0.7 0.66 1.31 ∗ 106 0.4883% 100.0% 4.531% 84.5% 83.3% 83.3% 7.1%

320.48 0.7 0.65 9.83 ∗ 105 0.3662% 75.0% 5.625% 85.2% 85.2% 84.8% 5.6%0.48 0.6 0.65 9.83 ∗ 105 0.3662% 75.0% 5.625% 85.2% 85.2% 84.8% 5.6%0.48 0.6 0.65 9.83 ∗ 105 0.3662% 75.0% 5.703% 85.5% 84.5% 83.8% 6.6%

PSML-CPCA-Components (PSML-CPCA-C)

640.46 0.6 0.70 1.31 ∗ 106 0.4883% 100.0% 3.828% 90.1% 89.8% 89.8% 0.6%0.46 0.6 0.70 1.31 ∗ 106 0.4883% 100.0% 3.984% 90.6% 90.0% 89.7% 0.7%0.46 0.6 0.70 1.31 ∗ 106 0.4883% 100.0% 3.828% 90.2% 89.9% 89.7% 0.7%

320.49 0.7 0.65 9.83 ∗ 105 0.3662% 75.0% 7.656% 87.0% 86.1% 85.9% 4.5%0.47 0.6 0.65 9.83 ∗ 105 0.3662% 75.0% 7.266% 86.8% 86.6% 85.8% 4.6%0.47 0.6 0.65 9.83 ∗ 105 0.3662% 75.0% 7.656% 86.6% 86.5% 85.8% 4.6%

PSML-CPCA-Mixed (PSML-CPCA-M)

640.46 0.6 0.65 1.31 ∗ 106 0.4883% 100.0% 3.672% 90.0% 90.0% 89.8% 0.6%0.46 0.6 0.65 1.31 ∗ 106 0.4883% 100.0% 3.750% 89.9% 89.8% 89.5% 0.9%0.46 0.6 0.65 1.31 ∗ 106 0.4883% 100.0% 3.594% 90.2% 89.7% 89.4% 1.0%

320.46 0.6 0.65 9.83 ∗ 105 0.3662% 75.0% 5.234% 88.9% 88.9% 88.8% 1.6%0.46 0.6 0.65 9.83 ∗ 105 0.3662% 75.0% 5.391% 88.8% 88.8% 88.4% 2.0%0.46 0.6 0.65 9.83 ∗ 105 0.3662% 75.0% 5.234% 89.0% 88.6% 88.3% 2.1%

* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table 9.25: Top three achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for each of thethree different template types with PSML-CPCA input at T = 64 and T = 32.

First, the performance of the three PSML-CPCA types are compared at T = 64 and

105

T = 32. The top three of every type and T combination is listed in table 9.25. Simply using theSMR-CPCA-A type of the PSML-CPCA, further called PSML-CPCA-A, fails to yield goodTP0, TP0.1 and TP0.5 due to a high overlap between genuine and impostor scores. Employ-ing both components (SMR-CPCA-C) of the PSML-CPCA, further called PSML-CPCA-C,drastically increases the biometric performance of the CPCA-Tree and reduces the PER by afraction. However, the hybrid type (SMR-CPCA-M) of the PSML-CPCA, further called PSML-CPCA-M, achieves a equivalent biometric performance as the PSML-CPCA-C for T = 64 butoutperforms it for T = 32 due to the reduced population in the sign-bit.Even in the most basic implementation, the CPCA-Tree is able to increase the biometric per-formance compared to the naïve binary PSML-CPCA approach without additional workloadreduction (T = 64; PSML-CPCA-M and PSML-CPCA-C), and nearly reaches the biometricperformance of the naïve binary PSML-CPCA with a workload reduction of 25% (T = 32;PSML-CPCA-M).

9.5.2 Tree pre-selection

By narrowing down the feasible types for T = 64 to the PSML-CPCA-C and PSML-CPCA-M,for T = 32 to PSML-CPCA-M, the next experiments are run to determine the workload reduc-tion capabilities with tree pre-selection. As in the Bloom filter sections, the tree pre-selectionexperiments for T = 64 are run with t = 1, . . . , 48 trees and t = 1, . . . , 26 trees for T = 32since selecting more trees would increase the workload.

Beginning with the PSML-CPCA-C, the tree pre-selection results are listed in table 9.26.The quick decay of the biometric performance is very conspicuous: already at a pre-selectionof t = 1/2 trees, the biometric performance loses 0.5%-points in TP0 while only achieving anadditional workload reduction of ∼13% compared to the naïve PSML-CPCA baseline.

It can be observed from table 9.27 that the PSML-CPCA-M CPCA-Tree performs muchbetter with enabled tree pre-selection. With as little TP0 loss as 0.2%-points compared to thenaïve binary SMR-CPCA approach, a workload reduction of ∼32% is achieved. A workloadreduction of ∼50% is achievable with a performance gain of ∼1%-point compared to the naïvebinary PSML-CPCA approach.These better results are trivially explainable with figure 7.3 and table 7.1. T = 64 yields 4 tem-plates per tree. Comparing four merged (joined using a binary OR-operation) sign-bits (figure7.3g) and four merged applied-bit16 (figure 7.3i), the merged sign-bit has an average popula-tion rate of 93.5% while the applied-bit has an average population rate of 61.4%. Therefore, theroot node of a CPCA-Tree using PSML-CPCA-C templates at T = 64 is nearly set completelyand thus does not contain any useful information. A tree pre-selection is therefore renderedserendipity, which may yield acceptable results with a high pre-selection rate, but drastically

16Which are identical to the sign-bit of the PSML-CPCA-M.

106

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.47 0.64 0.65 3.48 ∗ 105 0.1297% 26.6% 49.61% 49.3% 48.8% 48.6% 41.8%2 0.49 0.63 0.65 3.69 ∗ 105 0.1373% 28.1% 37.27% 61.0% 60.5% 60.4% 30.0%3 0.49 0.63 0.65 3.89 ∗ 105 0.1450% 29.7% 29.14% 68.9% 68.3% 68.1% 22.3%

. . . ‡

31 0.46 0.62 0.70 9.63 ∗ 105 0.3586% 73.4% 4.69% 90.1% 89.3% 89.3% 1.1%32 0.46 0.62 0.70 9.83 ∗ 105 0.3662% 75.0% 4.53% 90.2% 89.4% 89.4% 1.0%33 0.46 0.62 0.65 1.00 ∗ 106 0.3738% 76.6% 4.06% 89.8% 89.7% 89.5% 0.9%

. . . ‡

39 0.46 0.62 0.70 1.13 ∗ 106 0.4196% 85.9% 4.14% 89.9% 89.7% 89.7% 0.7%40 0.46 0.62 0.70 1.15 ∗ 106 0.4272% 87.5% 3.98% 90.0% 89.8% 89.8% 0.6%

. . . ‡

46 0.46 0.62 0.70 1.27 ∗ 106 0.4730% 96.9% 3.83% 90.0% 89.8% 89.8% 0.6%47 0.46 0.62 0.70 1.29 ∗ 106 0.4807% 98.4% 3.75% 90.1% 89.8% 89.8% 0.6%48 0.46 0.62 0.70 1.31 ∗ 106 0.4883% 100.0% 3.75% 90.1% 89.8% 89.8% 0.6%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.14.

Table 9.26: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using PSML-CPCA-C templates at T = 64.

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.64 0.65 3.48 ∗ 105 0.1297% 26.6% 28.75% 70.2% 69.5% 69.3% 21.1%2 0.49 0.64 0.65 3.69 ∗ 105 0.1373% 28.1% 18.44% 79.0% 78.4% 78.0% 12.4%3 0.49 0.67 0.70 3.89 ∗ 105 0.1450% 29.7% 14.22% 83.0% 82.7% 81.9% 8.5%

. . . ‡

16 0.47 0.61 0.65 6.55 ∗ 105 0.2441% 50.0% 5.08% 89.3% 89.2% 89.0% 1.4%. . . ‡

21 0.46 0.62 0.65 7.58 ∗ 105 0.2823% 57.8% 4.22% 89.8% 89.8% 89.5% 0.9%22 0.46 0.62 0.65 7.78 ∗ 105 0.2899% 59.4% 4.14% 89.8% 89.8% 89.6% 0.8%23 0.46 0.62 0.65 7.99 ∗ 105 0.2975% 60.9% 3.91% 89.9% 89.9% 89.7% 0.7%

. . . ‡

28 0.46 0.62 0.65 9.00 ∗ 105 0.3357% 68.7% 3.75% 90.0% 90.0% 89.8% 0.6%. . . ‡

48 0.46 0.62 0.65 1.31 ∗ 106 0.4883% 100.0% 3.67% 90.0% 90.0% 89.8% 0.6%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.15.

Table 9.27: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using PSML-CPCA-M templates at T = 64.

increases the PER when using small pre-selection rates.Looking at the results for T = 32 using the PSML-CPCA-C templates in table 9.28 a workload

107

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.60 0.70 1.95 ∗ 105 0.0725% 14.84% 55.23% 43.9% 43.8% 43.7% 46.7%2 0.49 0.60 0.75 2.25 ∗ 105 0.0839% 17.19% 41.25% 57.4% 57.3% 57.0% 33.4%3 0.49 0.60 0.75 2.56 ∗ 105 0.0954% 19.53% 33.20% 65.2% 64.5% 64.5% 25.9%

. . . ‡

20 0.46 0.62 0.65 7.78 ∗ 105 0.2899% 59.38% 7.11% 87.7% 87.7% 87.6% 2.8%21 0.46 0.62 0.65 8.09 ∗ 105 0.3014% 61.72% 6.72% 88.1% 88.1% 88.0% 2.4%22 0.46 0.62 0.65 8.40 ∗ 105 0.3128% 64.06% 6.48% 88.3% 88.3% 88.1% 2.3%23 0.46 0.62 0.65 8.70 ∗ 105 0.3242% 66.41% 6.40% 88.4% 88.4% 88.2% 2.2%24 0.46 0.62 0.66 9.01 ∗ 105 0.3357% 68.75% 6.02% 88.7% 88.7% 88.5% 1.9%25 0.46 0.62 0.66 9.32 ∗ 105 0.3471% 71.09% 5.94% 88.7% 88.7% 88.5% 1.9%26 0.46 0.62 0.66 9.63 ∗ 105 0.3586% 73.44% 5.62% 88.8% 88.8% 88.7% 1.7%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.16.

Table 9.28: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 26 trees using PSML-CPCA-M templates at T = 32.

reduction using lesser trees and lesser trees with tree pre-selection is not feasible: using T = 64with a small tree pre-selection yields a much better workload reduction with respect to the bio-metric performance. The CPCA-Tree with T = 32 using the PSML-CPCA-C templates suffersa high PER referable to the same problem as for the PSML-CPCA-C templates: a higher pop-ulation rate in the root node of every tree aggravates a successful tree pre-selection. Therefore,it is advisable to keep the population rate in the tree roots smaller than ∼75%, desirable at∼50% to ∼70%.

At this point, the PSML-CPCA-M results in an additional workload reduction of 50% com-pared to the naïve binary PSML-CPCA approach at the best biometric performance achievedwith the binary PSML-CPCA Bloom filter, respectively with the biometric performance of thenaïve binary PSML-CPCA approach. It is hardly expectable to achieve a higher workload re-duction with the CPCA-Tree than with the binary PSML-CPCA Bloom filter. However, lookingat the low PER of the CPCA-Tree it might be possible to increase the biometric performanceof the binary PSML-CPCA without increasing the workload compared to the naïve approach.The next experiments are run to achieve this goal.

9.5.3 Additional binary comparison

As already conducted with the Bloom filter approaches, an additional binary PSML-CPCAtemplate comparison could help to increase the biometric performance since impostor andgenuine scores could be more separated. However, using the additional comparison approach isexpected to only affect the PSML-CPCA-A templates since CPCA-Trees with PSML-CPCA-C

108

or PSML-CPCA-M templates already undertake these comparisons while traversing the trees.This expectation is proven in the appended table A.17 in comparison with table 9.27: using anadditional binary PSML-CPCA template comparison for the PSML-CPCA-M does not yieldany improved biometric performance but increases the workload by one comparison.

An experiment using the additional binary comparison for CPCA-Trees with PSML-CPCA-A templates is skipped. Due to the larger PER in CPCA-Trees with PSML-CPCA-A templates,it is not expected that an additional comparison will increase the biometric performance throughthe level of CPCA-Trees with PSML-CPCA-M templates.

9.5.4 Masking out most common bits

Masking out the most common bits in a binary PSML-CPCA template could increase the bio-metric performance since it is more difficult for impostor queries to be accepted while genuinequeries are ideally unaffected. Since the PSML-CPCA-M templates achieve the highest biomet-ric performance paired with the highest workload reduction, the PSML-CPCA-M templateswith T = 64 and T = 32 are used in this experiment.

To create a mask with the most common bits, a heat map is used: the set bits of all enrolledtemplates are summed up and all sums are normalised to [0, 1]. The mask is created by selectingall normalised sums exceeding a given threshold17.Masking out the most common bits successfully reduced the PER and increased TP0 by 0.1%-points for both T = 64 (table 9.29) and T = 32 (table 9.30). For T = 64, a bit threshold of 0.6to 0.7 seems feasible.

It has to be noted that the experiment can be further optimised. The heat map used forthis experiment was created using a fixed binarisation threshold of 0.7. Better results could beachieved when creating the heat map with every tested mask threshold.

9.5.5 Additional real-valued comparison

Employing an additional real-valued PSML-CPCA comparison noteably increased the biomet-ric performance for the Bloom filter approaches. Since the CPCA-Tree yield comperatively lowPER, it could also gain in TP0. The experiment is only run with 1, . . . , 32 pre-selected trees:selecting more trees would increase the workload due to the additional real-valued comparisonand the tree pre-selection workload.The TP0 values in tables 9.31 and 9.32 are disappointing, while the TP0.1 values indicate a

much higher performance. When analysing the TP values for every threshold at T = 64, it isnoticeable that the bad TP0 bases on one single impostor query that early exceeds the decision

17Called bit threshold in the tables.

109

SMR Workload Recognition Performance

t λmax λCP CAmax MT Bit Threshold ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.65 0.65 0.7 3.53 ∗ 105 0.1316% 26.9% 28.91% 70.0% 69.7% 69.3% 21.1%2 0.49 0.64 0.70 0.5 3.74 ∗ 105 0.1392% 28.5% 18.67% 79.1% 78.4% 78.4% 12.0%3 0.49 0.64 0.70 0.6 3.94 ∗ 105 0.1469% 30.1% 14.22% 82.9% 82.1% 82.0% 8.5%

. . . ‡

16 0.46 0.62 0.70 0.6 6.60 ∗ 105 0.2460% 50.4% 4.844% 89.5% 89.1% 89.1% 1.3%. . . ‡

21 0.46 0.62 0.65 0.8 7.63 ∗ 105 0.2842% 58.2% 4.219% 89.8% 89.8% 89.5% 0.9%22 0.46 0.62 0.65 0.7 7.83 ∗ 105 0.2918% 59.8% 4.219% 89.8% 89.8% 89.6% 0.8%23 0.46 0.62 0.65 0.7 8.04 ∗ 105 0.2995% 61.3% 3.984% 89.9% 89.9% 89.7% 0.7%

. . . ‡

28 0.46 0.62 0.70 0.6 9.06 ∗ 105 0.3376% 69.1% 3.828% 90.2% 89.9% 89.9% 0.5%. . . ‡

46 0.46 0.62 0.70 0.5 1.27 ∗ 106 0.4749% 97.3% 3.750% 90.2% 89.9% 89.9% 0.5%47 0.46 0.62 0.70 0.6 1.30 ∗ 106 0.4826% 98.8% 3.672% 90.2% 89.9% 89.9% 0.5%48 0.46 0.62 0.70 0.6 1.32 ∗ 106 0.4902% 100.4% 3.672% 90.2% 89.9% 89.9% 0.5%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.18.

Table 9.29: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using PSML-CPCA-M templates at T = 64 with an additional binaryPSML-CPCA comparison and masking out most common set bits.

SMR Workload Recognition Performance

t λmax λCP CAmax MT Bit Threshold ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.60 0.70 0.9 2.00 ∗ 105 0.0744% 15.23% 55.23% 43.9% 43.8% 43.7% 46.7%2 0.49 0.60 0.75 0.8 2.30 ∗ 105 0.0858% 17.58% 41.25% 57.5% 57.4% 57.0% 33.4%3 0.49 0.60 0.75 0.5 2.61 ∗ 105 0.0973% 19.92% 32.97% 65.4% 64.7% 64.7% 25.7%

. . . ‡

20 0.48 0.66 0.75 0.5 7.83 ∗ 105 0.2918% 59.77% 6.95% 88.3% 88.2% 87.7% 2.7%21 0.48 0.66 0.75 0.5 8.14 ∗ 105 0.3033% 62.11% 6.64% 88.6% 88.5% 88.0% 2.4%22 0.48 0.66 0.75 0.5 8.45 ∗ 105 0.3147% 64.45% 6.02% 89.1% 88.7% 88.2% 2.2%23 0.48 0.66 0.75 0.5 8.76 ∗ 105 0.3262% 66.80% 5.86% 89.1% 88.8% 88.3% 2.1%24 0.46 0.62 0.65 0.9 9.06 ∗ 105 0.3376% 69.14% 6.02% 88.7% 88.7% 88.5% 1.9%25 0.48 0.66 0.75 0.5 9.37 ∗ 105 0.3490% 71.48% 5.86% 89.1% 89.1% 88.6% 1.8%26 0.46 0.62 0.65 0.7 9.68 ∗ 105 0.3605% 73.83% 5.70% 88.8% 88.8% 88.7% 1.7%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.19.

Table 9.30: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 26 trees using PSML-CPCA-M templates at T = 32 with an additional binaryPSML-CPCA comparison and masking out most common set bits.

threshold. If this query is removed from the query set, TP0 = 90.8%18 is achieved when pre-selecting 32 trees and TP0 = 90.3% when pre-selecting 16 trees. However, in terms of TP0, the

18∼0.4%-points more than the naïve real-valued PSML approach; does not increase the workload comparedto the naïve binary PSML-CPCA approach while achieving a much higher biometric performance.

110

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.64 0.65 6.76 ∗ 105 0.2518% 51.56% 28.67% 70.5% 70.2% 69.8% 20.6%2 0.49 0.64 0.70 6.96 ∗ 105 0.2594% 53.12% 17.81% 80.5% 80.1% 79.9% 10.5%3 0.49 0.64 0.70 7.17 ∗ 105 0.2670% 54.69% 13.98% 84.1% 83.8% 83.6% 6.8%

. . . ‡

9 0.47 0.65 0.70 8.40 ∗ 105 0.3128% 64.06% 6.094% 89.6% 89.4% 88.8% 1.7%. . . ‡

14 0.48 0.66 0.70 9.42 ∗ 105 0.3510% 71.88% 4.609% 90.6% 90.2% 88.0% 2.4%15 0.48 0.65 0.70 9.63 ∗ 105 0.3586% 73.44% 4.219% 90.9% 90.4% 88.1% 2.3%16 0.48 0.65 0.75 9.83 ∗ 105 0.3662% 75.00% 4.062% 90.8% 90.3% 88.1% 2.3%

. . . ‡

30 0.48 0.64 0.65 1.27 ∗ 106 0.4730% 96.88% 3.359% 91.2% 90.9% 88.5% 1.9%31 0.47 0.64 0.70 1.29 ∗ 106 0.4807% 98.44% 3.359% 91.2% 90.8% 88.5% 1.9%32 0.47 0.64 0.70 1.31 ∗ 106 0.4883% 100.00% 3.359% 91.2% 90.8% 88.5% 1.9%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.20.

Table 9.31: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 32 trees using PSML-CPCA-M templates at T = 64 with an additional realPSML-CPCA comparison.

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.60 0.70 5.22 ∗ 105 0.1945% 39.84% 55.23% 44.1% 43.8% 43.8% 46.7%2 0.49 0.60 0.75 5.53 ∗ 105 0.2060% 42.19% 41.25% 58.0% 57.7% 57.3% 33.1%3 0.48 0.65 0.75 5.84 ∗ 105 0.2174% 44.53% 32.66% 66.0% 65.9% 65.6% 24.8%

. . . ‡

21 0.47 0.64 0.70 1.14 ∗ 106 0.4234% 86.72% 5.78% 89.6% 89.1% 87% 3.5%22 0.47 0.64 0.70 1.17 ∗ 106 0.4349% 89.06% 5.62% 89.7% 89.2% 87% 3.5%23 0.48 0.62 0.70 1.20 ∗ 106 0.4463% 91.41% 5.31% 89.5% 89.2% 88.5% 1.9%24 0.49 0.65 0.65 1.23 ∗ 106 0.4578% 93.75% 5.08% 89.8% 89.5% 87.3% 1.1%25 0.49 0.65 0.65 1.26 ∗ 106 0.4692% 96.09% 4.92% 90% 89.6% 87.5% 2.9%26 0.48 0.62 0.65 1.29 ∗ 106 0.4807% 98.44% 4.77% 90.2% 89.8% 87.5% 2.9%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.21.

Table 9.32: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 26 trees using PSML-CPCA-M templates at T = 32 with an additional realPSML-CPCA comparison.

approach using an additional binary comparison yields a higher performance. It has to be notedthat the configuration using a pre-selection of 15 trees in table 9.31 surpasses all naïve binaryworkload-reduction-by-feature-reduction approaches while introducing an additional workloadreduction of ∼27% compared to the naïve binary PSML-CPCA approach. The highlighted row

111

in table 9.31 achieves the highest TP0 with just 9 searched trees. Since this is the only con-figuration in this area and TP0.1 and TP0.5 are lower than the configuration with 10 searchedtrees, it could be serendipity.

Another experiment with most common bit masking and additional real comparison wasrun but did not achieve better results. Refer to appended table A.22 for the results.

9.6 Using the full vascular pattern in the SMRFacing the overall acceptable baseline performance in intra-session experiments but poor baselineperformance in the inter-session experiments, an additional experiment has been run to exam-ine whether the SMR hits its limit in this fuzzy environment, or if employing more informationachieves a higher biometric performance.The average count19 of minutiae in the PolyU dataset is ∼70, whilst the whole vascular patternis represented by an average of ∼950 points. Therefore, ∼93% of the extracted information isnot represented in the SMR. Using more information does not necessarily result in an increasedbiometric performance, although the chances are high that with a careful selection concerningwhich parts of the vein pattern to include as input for the SMR, at least the overlap betweenimpostor and genuine attempt scores shrinks.The results in figure 9.20 show a significant biometric performance improvement over the whole

10−3 10−2 10−1 100

0.1

0.2

0.3

0.4

0.50.5

0.6

0.7

0.8

0.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

Minutiae Inter-Session (Absolute)Minutiae Inter-Session (Real)Pattern Inter-Session (Absolute)Pattern Inter-Session (Real)

(a) inter-session

10−3 10−2 10−1 100

0.1

0.2

0.3

0.4

0.50.5

0.6

0.7

0.8

0.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

Minutiae Intra-Session (Absolute)Minutiae Intra-Session (Real)Pattern Intra-Session (Absolute)Pattern Intra-Session (Real)

(b) intra-session

Figure 9.20: ROC of the PSML-Baseline (minutiae PSML) and full vein pattern SML.

ROC for inter-session and intra-session data. Table 9.33 lists the corresponding TP values forfigure 9.20. Note how the full vascular pattern SML in the inter-session experiment achievesa similar or higher biometric performance than the intra-session experiments using minutiaePSML.

19Before the removal of spurious minutiae.

112

Experiment Recognition Performance

Session Input TP0.5 TP0.1 TP0 EER

Inter Pattern 92.9% 91.0% 90.5% 1.3%Intra Pattern 97.9% 97.7% 97.0% 0.28%Inter Minutiae 78.1% 75.1% 72.4% 3.4%Intra Minutiae 91.2% 91.1% 90.4% 2.1%

Table 9.33: Biometric performance comparison between minutiae PSML and vascular pattern SML.

Therefore, using the full vascular pattern for the SMR greatly increases the biometric per-formance without increasing the workload in any means: the full vascular pattern SMR issampled in the same way and size as the minutiae SMR and the same workload reductionapproaches can be utilised. Only the off-line cost of calculating the SMR is increased.

9.7 SummaryIn summary, only five workload reduction methods, two using Bloom filter-indexing and threeusing CPCA-Tree-indexing, are of concern and listed in table 9.34.

Workload Recognition Performance

Name ω256 𝟋B* TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

PSML-Baseline† 2.68 ∗ 108 − 91.2% 91.1% 90.4% −Binary PSML-CPCA† 1.31 ∗ 106 0.4880% 89.7% 89.1% 89.0% 1.4%Bf CPCA ABC 9.22 ∗ 105 0.3437% 89.5% 88.9% 88.0% 2.4%Bf CPCA ARC 5.23 ∗ 105 0.1952% 89.9% 89.1% 89.0% 1.4%C-T SCM 9.01 ∗ 105 0.3357% 90.0% 90.0% 89.8% 0.6%C-T SCM Masking 9.06 ∗ 105 0.3376% 90.2% 89.9% 89.9% 0.5%C-T SCM ARC 1.13 ∗ 106 0.4196% 91.2% 90.8% 88.5% 1.9%* 𝟋 compared to the PSML ω256.† Naïve approach.

Table 9.34: Summary of relevant experiments, respectively workload reduction methods, with theirworkload and biometric performance.

The configurations listed in the table are discussed below.

Bf CPCA ABC Using the Bloom filter with the binary PSML-CPCA and an additionalbinary comparison achieved, as expected due to the usage of the PSML-CPCA, a muchbetter workload reduction, but fails to reach the biometric performance of the binaryPSML Bloom filter with additional binary comparison. The best workload-to-biometric-

113

performance trade-off is achieved with a Bloom filter size of 5×5 with a MT of 0.6 att = 19, 20, 21, yielding ω256 ≈ 9.22 ∗ 106 (𝟋B = 0.3437%) at TP0 = 88.0%.

Bf CPCA ARC To address the biometric performance loss of the Bloom filter with binaryPSML-CPCA and an additional binary comparison, the additional binary comparison isreplaced with an additional real-valued PSML-CPCA comparison (ARC). This methodcombines a low workload with a high biometric performance. Due to a more workload-reducing Bloom filter size of 6×4, the even smaller workload of 5.23∗105 (𝟋B = 0.1952%)at a biometric performance of TP0 = 89.0% with t = 22 is achieved. This is the recom-mended Bloom filter approach.

C-T SCM The CPCA-Tree-indexing (C-T) approach using the PSML-CPCA-M (SCM) tem-plates, which proved to achieve the best results in all CPCA-Tree configurations, alreadyachieves the highest biometric performance in the most basic implementation and is ableto maintain the high biometric performance with a workload of ∼70% to the naïve bin-ary PSML-CPCA. It has to be noted that the biometric performance is above this naïveapproach. The highest TP0 of 89.8% is reached at t = 28 with 𝟋B = 0.3357%, whichis ∼30% workload less than the binary PSML-CPCA at a ∼0.8%-points higher TP0. Ifthe biometric performance of a naïve binary PSML-CPCA is desired, the workload canfurther be reduced to 𝟋B = 0.2441 (half of the naïve approach) with t = 16.

C-T SCM Masking Masking out the most common bits set in a PSML-CPCA-M templatefor CPCA-Tree-indexing increased the maximum TP0 to 89.9% for the sake of one ad-ditional binary comparison20. A TP0 of 89.9% is reached at t = 28 with 𝟋B = 0.3376%(TP0 = 89.8% is achieved with 𝟋B = 0.3147%). Again, if the biometric performanceof a naïve binary PSML-CPCA is desired, the workload can further be reduced to𝟋B = 0.2384% with t = 15.

C-T SCM ARC Employing an additional real comparison to the CPCA-Tree with PSML-CPCA-M templates increased both TP0.1 and TP0.5 to the highest values of all workloadreduction experiments, but reduces TP0. Therefore, this configuration is not recommen-ded for high security scenarios. When selecting t = 23, the biometric performance ofTP0 = 88.5% and TP0 = 90.8% is achieved with at 𝟋B = 0.4196%.

Even if the Bloom filter-indexing using the binary PSML achieved an acceptable biometricperformance, it is not able to reach the workload reduction achieved with the indexing meth-ods using the PSML-CPCA. Since the CPCA-Tree-indexing methods yield a similar or betterbiometric performance, paired with an much lower workload, the binary PSML Bloom filter-indexing is not considered necessary to highlight in this summary.

20When comparing equal TP0 values for the two approaches, the workload of the masking approach is actuallylower.

114

For a better understanding of the workload-to-biometric-performance trade-off, figure 9.21shows the workload reduction achieved in relation to the achieved biometric performance. Itis clearly visible that the CPCA-Tree-indexing with and without most common bits mask-ing achieves a better biometric performance up to 𝟋B≈0.23% in the means of TP0. If a lesserworkload is desired, the PSML-CPCA Bloom filter with an additional real-valued PSML-CPCAcomparison passes the CPCA-Tree.

To deterministically benchmark the different indexing methods, the euclidean distancebetween an optimal operation point (OT P = 90.4%, O𝟋 = 0.1%) and each performance-to-workload sample from each experiment as shown in equation 9.1 can be used.

δ(TP,𝟋B) =√

(TP −OT P )2 + (𝟋B −O𝟋)2 (9.1)

The smallest δ(TP,𝟋B) for each experiment rates the corresponding indexing method. Choosingthe baseline biometric performance (instead of TP0 = 100%) as optimal operation point movesthe emphasis more to the workload reduction than the biometric performance, to even thescaling differences (TP0 scales from ∼70% to ∼90% while 𝟋B scales between ∼0.15% and∼1.0%).

Name δ(TP0,𝟋B) 𝟋B* TP0

C-T SCM Masking 0.55 0.3376 89.9C-T SCM 0.64 0.3357 89.8Bf CPCA ARC 1.40 0.1952 89.0Binary PSML-CPCA† 1.45 0.4880 89.0C-T SCM ARC 1.61 0.3128 88.8Bf CPCA ABC 2.41 0.3437 88.0

Name δ(TP0.1,𝟋B) 𝟋B* TP0.1

C-T SCM ARC 0.40 0.4425 90.9C-T SCM 1.12 0.3357 90.0C-T SCM Masking 1.22 0.2995 89.9Bf CPCA ARC 1.31 0.2155 89.8Binary PSML-CPCA† 2.04 0.4880 89.0Bf CPCA ABC 2.21 0.3437 88.9* 𝟋 compared to the PSML ω256.† Naïve approach.

Table 9.35: Ranking for each relevant experiment, respectively workload reduction method, ordered byδ(TP,𝟋B) with the optimal operation point at PSML-Baseline biometric performance (TP0 = 90.4)and 𝟋 = 0.1%.

Table 9.35 lists the minimum recorded δ(TP,𝟋B) for every indexing method and figure

115

9.22 plots their ROC. For both TP0 and TP0.1, the CPCA-Tree-indexing methods clearly yieldthe smallest δ(TP,𝟋B) since the CPCA-Tree achieves the highest TP0 at an average workloadreduction. When moving the optimal operation point to the biometric performance of the naïvebinary PSML-CPCA approach and 𝟋 = 0.1%, the Bloom filter-indexing using binary PSML-CPCA templates with an additional real-valued PSML-CPCA comparison takes the lead aslisted in appended table A.33.

Finally observe the figures in 9.21, for nearly every indexing method, the achieved TP valuesrise with decreasing t until t≈3/5. The reduced number of template comparisons needed reducesthe number of false matches as predicted by the at-least-one-false-match problem described insection 2.2.3.

The relevant experiments listed in this summary were repeated with inter-session data fromthe PolyU dataset and also repeated only for the second session of the dataset. They confirmmost general statements by showing the same behaviour as the main experiments listed hererun on the first session of the dataset. However, some differences between the intra-session andinter-session experiments are recorded and listed below.

• The Bloom filter-indexing methods suffer in biometric performance from the increasednumber of bit-errors in the binary SMR and SMR-CPCA templates in inter-session exper-iments. While the Bloom filter-indexing with additional real-valued SMR-CPCA compar-ison yields a TP0 performance loss of ∼1.5% for 𝟋B ≈ 0.2% in intra-session experiments,a TP0 performance loss of ∼5% for 𝟋B ≈ 0.2% is recorded in inter-session data.

• For intra-session data, masking out the most common bits in CPCA-Tree-indexing yieldeda TP0 biometric performance gain of ∼0.1%, for inter-session data a TP0 performancegain up to ∼1% is recorded.

• While in intra-session experiments using an additional real-valued SMR-CPCA compar-ison reduced the TP0 performance for CPCA-Tree-indexing, using an additional real-valued SMR-CPCA comparison in inter-session experiments increased the TP0 perform-ance by ∼0.7%. However, masking out the most common bits yields a higher TP0 per-formance.

116

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.869

70

72

74

76

78

80

82

84

86

88

90

𝟋B

TP

0%

Naïve binary PSML-CPCAOptimal operation pointBf CPCA ABCBf CPCA ARCC-T SCMC-T SCM MaskingC-T SCM ARC

(a) TP0

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.869

70

72

74

76

78

80

82

84

86

88

90

𝟋B

TP

0.1

%

Naïve binary PSML-CPCAOptimal operation pointBf CPCA ABCBf CPCA ARCC-T SCMC-T SCM MaskingC-T SCM ARC

(b) TP0.1

Figure 9.21: TP0 and TP0.1 performance in respect to 𝟋B values for relevant experiments at selectedt. Used abbreviations: Bf (Bloom filter), C-T (CPCA-Tree), ABC (Additional Binary Comparison),ARC (Additional Real Comparison), SCM (PSML-CPCA-M).

117

10−3 10−2 10−1 1000.87

0.89

0.91

0.93

0.95

0.96

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

PSML-BaselineBf CPCA ABCBf CPCA ARCC-T SCMC-T SCM ARCC-T SCM Masking

Figure 9.22: ROC for the with δ(TP,𝟋B) selected configurations.

118

9.8 ConclusionsThe SMR is a feasible representation for a efficient biometric identification. 75 000 binarySMR comparisons per second, ∼0.013ms per comparison, are achieved in a single-threadedexperiment on a modern Intel i7 CPU. Nearly 300 000 comparisons per second, ∼ 0.003ms

per comparison, are achieved using the binary SMR-CPCA. With the workload reduction ap-proaches used, the comparison speed is not necessarily increased21 but the amount of comparedtemplates is reduced. Using e.g. the CPCA-Tree-indexing with t≈3/8 reduces the amount ofcompared templates by around ∼40% without a significant loss in biometric performance, thusin a S = 100 000 identification scenario, only ∼60 000 template comparisons are needed.

However, the baseline biometric performance achieved with the PSML only using extractedminutiae, especially in inter-session scenarios, is not feasible for a real-world scenario. The lastexperiment presented in section 9.6 shows that the biometric performance can be increasedto real-world capable levels by employing more data in the SMR then only bifurcations orendpoints of the vascular pattern. With an increased biometric performance, it might be possibleto achieve a lower workload.

21Only in some Bloom filter configurations an increased comparison speed is expected, since some configura-tions result in a smaller representation than the given input.

119

Chapter 10

Discussion

In this chapter, the achieved results of the experiments in the previous chapter are discussed.Additionally, several future research possibilities for palm vein indexing are presented.

10.1 ResultsThis section discusses issues that occurred during the course of this project and several obser-vations that did not fit in the result chapter.

10.1.1 Poor performance on the CASIA and PUT datasets

The results obtained for the CASIA and PUT datasets are unusable. Already by comparing theresults yielded by the MHD for the CASIA and PUT datasets with the results for the PolyU itis noticeable that there are issues with the feature extraction pipeline.In the early stages of this thesis, during the implementation of the feature extraction pipelineonly the PolyU dataset was available. Therefore, every algorithm was tested and tuned with thePolyU dataset. Tuning the first stages of the pipeline (e.g. the image enhancement, section 4.3)is a notoriously difficult task that requires much time1. Reconfiguration of the first stages alsorequires a reconfiguration of the following stages.In order to avoid placing the main objective (workload reduction experiments) at risk, no furtherreconfiguration of the pipeline for the datasets has been undertaken.

10.1.2 Towards SMR on palm-print datasets

As already mentioned multiple times, the PolyU and the CASIA (e.g. see section 8.1) data-sets aim for palm-print modalities. In a real-world application, the raw data acquired by theimage-capturing device contains different properties and the image-capturing device utilisesapproaches as described in section 4.1. Compare the two example images in figure 10.1: the left

1Even for the PolyU dataset it should be possible to achieve a higher performance when more time is spenton tuning the first stages.

120

(a) (b)

Figure 10.1: Raw data received by a (a) proprietary device and (b) by the PolyU dataset.

image is taken with an proprietary device that utilises the listed recommendations; the ROI isextracted using the introduced method. Compared to the PolyU dataset, it is notable that theproprietary device does not capture any skin texture and generates a very clean image of thepalm veins. Further, the ROI acquired in the PolyU dataset seems to be very small comparedto the ROI extracted from the proprietary device, thus capturing less veins.The characteristics of the PolyU dataset therefore does not meet some expectations of the SMR:

• The SMR is a minutiae-based approach and therefore does not benefit from the skintexture captured in the PolyU dataset: skin texture is recognized as spurious minutiae.

• The small ROI extracted in the PolyU contains much less minutiae for the SMR: onaverage, the PolyU dataset contains 70 minutiae per sample; the proprietary device dataleads to around 130 minutiae per sample.

Therefore, the PolyU dataset is not a ideal dataset for a palm vein recognition experiments.However, the PolyU dataset is the largest publicly-available dataset to the author’s knowledge;it is the only dataset that features enough subjects to enrol a sufficient number of templates inthe trees of the Bloom filter or CPCA-Tree indexing approaches while allowing a satisfactorynumber of impostor comparisons.

10.1.3 SML minutiae reliability quality data impact

Contrary to all expectations, the extracted quality data did not increase the biometric per-formance in the experiments. On further consideration of this behaviour, it appears that thequality data mainly affects absolute-valued SMR representations. Again, observe figure 9.3,note how both SML and QSML yield a nearly equal ROC in the real-valued representation.When comparing the absolute-valued representations of both SML and QSML, a strong dif-ference is visible: the quality enhanced QSML achieves a higher ROC, since the complex part,

121

which the absolute-valued representation is based on, is much more sensitive regarding singleminutiae.On the other hand, the SMC suffers from using the quality data (QSMC) in both representa-tions. This most likely is based on the higher sensitivity of the SMC due to the employmentof rotation information. The extracted rotation information is much more fuzzy than the loc-ation information. Using many minutiae, all of the same value, balances out some fuzziness.When reducing the value of some minutiae with reliability quality information, the other minu-tiae become much more dominant. If these dominant minutiae contain many errors in theirrotation-representation, resulting (mated) SMC also contain many errors.Therefore, in a fuzzy environment, the quality data should only be applied to the SML.

10.1.4 Ageing and environmental effects in the datasets

In section 9.2.4, the different results for same day (first session; intra-session) and differentday (first versus second session; inter-session) are presented. The doubling of the FNIR corres-ponds to the doubling of the FNMR reported by the International Biometric Group in [The06]between same day and different day results. Observe figure 10.2 where the DET curves of dif-ferent commercial biometric systems are shown. Note that the FNMR has doubled betweensame and different day experiments for the Fujitsu PalmSecure palm vein system as listed intable 10.1. Since DET and ROC curves are hardly comparable, figure 10.3 shows the same (first

Security* Same-Day/Intra-Session Different-Day/Inter-Session

FNMR FMR FNMR FMR

Low 3.13% 0.0380% 6.17% 0.0395%Default 4.23% 0.0118% 8.52% 0.0135%High 5.64% 0.0018% 11.86% 0.0007%* According to Fujitsu.

Table 10.1: FMR and FNMR for the Fujitsu PalmSecure palm vein biometric system taken from[The06].

session) and different day (first and second session) DET curve achieved with the PolyU datasetin an early state of this project where the pipeline was optimized for verification. With FNMR= 6.5% at FMR = 0.01% (same day) and FNMR = 13.6% at FMR = 0.01% (different day),the FNMR is doubled between same and different day, as expected from the results in [The06].

The main focus of this project is the workload reduction, not a competitive feature ex-traction pipeline. For this focus, it is irrelevant whether the data is generated by using oneor two sessions. Therefore, the exhaustive experiments only use the first session to better in-spect the workload reduction performance, not the feature extraction pipeline performance.However, with the results of FNMR = 6.5% at FMR = 0.01% by only using minutiae (not

122

Copyright © 2006 International Biometric Group

September 2006 CBT Round 6 Public Report International Biometric Group Executive Summary – 8

Figure 10.2: DET curves of different commercial biometric systems for same and different days (takenfrom [The06]).

10−6 10−5 10−4 10−3 10−2 10−1 10010−4

10−3

10−2

10−1

100

False Match Rate

False

Non

-Mat

chR

ate

Same-Day/Intra-SessionDifferent-Day/Inter-Session

Figure 10.3: DET curve for same (intra-session) and different day (inter-session) verification using thePolyU dataset achieved with the SMR system.

123

the whole palm vein structure and no other features), the pipeline yields acceptable resultscompared to the presented commercial product. Further, using the full vascular pattern (likethe commercial product), the proposed system outperforms the commercial product in intra-session experiments (FNMR = 1.28% at FMR = 0.01%) and yields competitive results forinter-session experiments (FNMR = 8.65% at FMR = 0.01%).

10.1.5 Small number of templates per Bloom filter tree

The reader may have noticed that using T = 64 yields only 4 templates per tree when 256templates are enrolled. This results in 4 comparisons per tree, therefore no workload reductionis achieved in terms of comparisons. The results showed that when reducing the tree count toT = 32 (8 templates, 6 comparisons per tree) or even T = 16 (16 templates, 8 comparisons pertree) the biometric performance suffers quickly and thus the approach is not feasible.In both Bloom filter-indexing sections, it is shown that nearly every column per block suffersfrom bit-errors. This forces very small Bloom filter block heights, whereby the bit errors appearin less blocks. However, reducing the block height automatically introduces more bit collisions,since the Bloom filter features less bits. Therefore, already comparatively similar templatesbecome even more similar in the Bloom filter representation.While this behaviour is desired in the verification mode, it is a huge disadvantage for theidentification scenario: similar enrolled templates are already also more similar in the Bloomfilter trees, thus it is more difficult to determine correct traversal direction or tree pre-selectiondecisions which increases the template pre-selection error. Further, with fewer Bloom filter bits,the tree root and all nodes (not leafs) are populated too quickly, which accelerates the pointwhere incorrect traversal direction or pre-selection decisions happen.

10.1.6 Results compared to state-of-the-art approaches

Different approaches based on local invariant feature extraction algorithms (i.e. [PK11, LT15])use the whole image for recognition tasks. Even if the publications state conducting a palmvein recognition, both actually use a palm print and palm vein combination.In [LT15], the verification results for palm print and palm vein are reported separately, albeitthe palm vein approach still uses the whole palm image. The authors tested their method onlyon 100 subjects with 1 reference and 5 probes per subject, with all samples taken from onesession. The in [LT15] proposed method achieved a EER of 2.4%, a different implementationusing the SURF algorithm achieved 3.1% and an implementation of the local binary patterns(LBP) approach achieved 0.7%. Table 10.2 compares the EER2 achieved in this project withthe reported biometric performance from the publication. Using the minutiae SMR achieves alower EER than the vascular pattern SURF and GSM approaches but fails to achieve a lowerEER than the LBP. With the full vascular pattern SMR, an EER of 0.28% is achieved. This low

2Recall, optimisation was done for identification, not for verification.

124

Input Method EER Verification (ms)§ Source

Minutiae SMR 2.11% 0.02 *

Minutiae Binary SMR 2.23% 0.01 *

Minutiae SMR-CPCA 2.22% 0.005 *

Minutiae Binary SMR-CPCA 2.24% 0.003 *

Pattern SMR 0.28% 0.02 *

Pattern Binary SMR 0.46% 0.01 *

Pattern SMR-CPCA 0.32% 0.005 *

Pattern Binary SMR-CPCA 0.52% 0.003 *

Pattern† LBP 0.71% 1.5 [LT15]Pattern† SURF 3.14% 88 [LT15]Pattern† GSM‡ 2.39% 1.7 [LT15]Pattern & Palm Print LBP 0.42% 1.5 [LT15]Pattern & Palm Print SURF 0.43% 88.0 [LT15]Pattern & Palm Print GSM‡ 0.30% 1.7 [LT15]* Experiments in this thesis.† Extended with some palm print structure.‡ Gray Surface Matching§ Specifications of machine used in [LT15] unknown; specifications of tested machine: 2014 Intel i7@ 3.60 GHz, single-threaded.

Table 10.2: Recorded EER of several intra-session experiments to compare the approaches in thisthesis with [LT15].

EER outperforms all other methods, even the vascular pattern and palm print combinations,and is much faster. The binary full vascular pattern SMR drops back to a EER of 0.46%, stilloutperforms the vascular pattern LBP, SURF and GSM, but is placed behind the vascularpattern and palm print combinations.

Another mentionable approach is presented by Abbas and George in [AG14]. This approachdiffers from the previous since it actually only observes the vein pattern as a biometric char-acteristic. Unfortunately, the publication does not offer an ISO/IEC-conforming performancereporting for the identification in a closed-set scenario and only reports the Correct RecognitionRate (CRR); the ratio of the number of samples being correctly classified to the total number oftested samples according to the publication. Further, is it not known which images of the PolyUdataset are taken as enrolled templates and whether the identification was carried out inter- orintra-session. Therefore, it is difficult to compare these results with those obtained through thisproject since no CRR optimisation experiments have been run and every identification experi-ment is run as an open-set. The publication also summarizes other vein recognition approachesfor different vein modalities not run on the PolyU dataset.However, to compare the results achieved in this thesis, experiments have been run to determine

125

the CRR. Without optimisation to achieve the highest CRR, the inter-session vascular pattern

Data Type Recognition Performance

Input Session Method Workload-Reduction CRR EER Identification (ms)‡ Source

Pattern Inter LAVD† - 99.95% 0.24% 158.4 [AG14]Pattern Inter SMR - 97.7% 1.28% 11.1 *

Pattern Inter SMR-CPCA - 97.7% 1.27% 1.2 *

Pattern Intra SMR - 99.8% 0.28% 11.1 *

Pattern Intra SMR-CPCA - 99.8% 0.32% 1.2 *

Minutiae Inter SMR - 93.7% 3.3% 11.1 *

Minutiae Intra SMR - 94.8% 2.1% 11.1 *

Minutiae Intra SMR-CPCA CPCA-Tree-indexing 94.7% 2.1% 0.9 *

* Experiments in this thesis.† Local Average of Vein Direction‡ Specifications of machine used in [AG14] unknown; specifications of tested machine: 2014 Intel i7 @ 3.60 GHz, single-threaded;identification at S = 500.

Table 10.3: Recorded CRR of several experiments to compare the approaches in this thesis with [AG14].

SMR experiment recorded a CRR of 97.7% which did not outperform the Local Average ofVein Direction (LAVD) method in terms of biometric performance. Comparing the executiontime of one single identification attempt, the SMR is more than 10 times, and the SMR-CPCAmore then 100 times faster than the LAVD. Again, the minutiae SMR are not able to achievegood results compared to the vascular pattern SMR and LAVD.

The results achieved with the vascular pattern SMR were quick experiments without op-timisation. Hence, it might be possible to achieve a even higher biometric performance withdedicated research.

10.2 Workload of real-valued SMR templatesRecall (section 9.3.1), the workload of one real-valued SMR (respective SMR-CPCA) compar-ison is calculated as 32 times the cost of a binary comparison as a floating point contains 32 bits.An experiment has been run to review this claim, since the measurements in section 10.1.6 hintat a workload increase of only ∼30%. 3 000 real-valued SMR (singe precision float) templatesand their binary representations3 were compared 1 000 times on a dedicated, completely strippeddown Linux server without graphical or network components. The server features a Intel Corei7-4790 CPU at 3.60 GHz with 8192 kB cache and is equipped with 16 GB RAM. To receivethe most clean results, the kernel and all remaining daemons were pinned to core 0, while theexperiment was executed on core 2. As a reference, a completely stock MacBook Air (Mid 2011,Intel Core i7-2677M at 1.80 GHz, 4 GB RAM) was used to verify the results.

3The binary representation is stored as 128×4 32 bit unsigned integers.

126

The experiment was repeated multiple times. On the Linux server, the real-valued compar-ison was only 1.3 times slower; the MacBook Air averaged 1.6 times slower.

Both processors are comparatively old (2011 and 2014). The newer processor (i7-4790)appears to have a much better optimised floating point pipeline, since it is relative floatingpoint speed is much higher than the older i7-2677M. There is no guarantee that these resultsalso apply to other architectures (ARM, PowerPC, …).It is safe to say, that a real-valued comparison does not yield a workload 32 times higherthan the binary approach, but an approximately 2 times higher workload for more realisticcalculations is reasonable. However, this can not be deterministically proven or generalisedfor all architectures or processors of the same architecture. Therefore the naïve approach ofcalculating the bit-count for each template is used in the results even if this does not reflectreal world scenarios or implementations.

10.3 Derived further research topicsAs already mentioned in section 10.1.6, during the course of this project several other researchtopics and ideas emerged that are not addressed in this thesis. This section lists the mostpromising derived research topics selected by the author, divided into one subsection for otherbiometric modalities and another for (palm) vein modalities.

10.3.1 (Palm) vein modalities

At least three further research topics for palm vein indexing can be derived from the resultsand other publications. Two research topics extend or modify the approach presented in thisthesis, whereas the third discards the whole SMR approach.

Employing more reference points. It was shown in section 9.6 that using the full vascularpattern as an input for the SMR strengthens the biometric performance in any concern.Without any optimisations, quality assurance or adjustments in any means, an inter-session TP0 = 90.5% (EER = 1.3%) and an intra-session TP0 = 97.0% (EER = 0.28%)was reached. Therefore, without any optimisations, an increase in TP0 of ∼18% and ∼7%,compared to the optimised results only using endpoints and bifurcations, is achieved. Itcan be safely assumed that with careful optimisations and exhaustive research muchhigher TP0 can be realised, which opens new possible avenues of research in the area.

Usage of absolute- and real-valued SMR in the CPCA-Tree. In the full vascular pat-tern experiments, the absolute-valued SMR achieved much higher CRR than the real-valued SMR, while the real-valued SMR achieved much higher TP0 values. This beha-viour could be used in e.g. the CPCA-Tree-indexing by building the CPCA-Tree withabsolute-valued binary SMR-CPCA to achieve a low PER and employ a final real-valuedSMR-CPCA comparison for a high TP0. Thus, the advantages of both SMR are combined,

127

which is expected to increase the overall biometric performance. Using both real-valuedand absolute-valued SMR is a non-issue in the means of computional workload since bothcan be extracted from one single SMR calculation.

Adoption of a different feature extraction approach. In [AG14], a simpler and seem-ingly more robust feature extraction approach is presented. A first next step could beto reimplement the proposed feature extraction approach and replace the current featureextraction pipeline before the thinning step (see section 4.5). This could be sufficient toincrease the overall biometric performance since it may yield more reliable minutiae.The presented feature extraction pipeline could still be used to extract quality informa-tion.

Bloom filter applied upon a not thinned vascular pattern. A more trivial and less com-plex approach could be to apply the Bloom filter template transformation to a segmented,binary vascular pattern. As an example, applying the Bloom filter on the 7th image offigure 3 in [AG14] with some additional dilation or other enhancements could alreadyachieve acceptable results.

Creation of large-scale palm vein dataset For obvious reasons, the creation of a dedic-ated, large-scale dataset for palm vein research is recommended. The desireable numberof subjects is S ≥ 1000 to cover both workload reduction and identification research.

10.3.2 Other modalities

The SMR is a versatile approach that can be used on most minutiae or reference-point basedapproaches in different modalities. These presented further research topics aim to extend andverify the presented indexing approaches based on the SMR in other modalities.

Fingerprint SMR-CPCA and CPCA-Trees. The feature representation used in this pro-ject is extrapolated from and originally designed for the fingerprint modality. While theSMR has not achieved a very high biometric performance using minutiae for the palmvein characteristic, it originally achieved a very high biometric performance for the minu-tiae of the fingerprint characteristic. This thesis has shown that the Bloom filter andCPCA-Tree indexing approaches can be applied to the SMR to reduce the overall work-load without losing a significant amount of biometric performance in this low-quality4.Therefore it would be a reasonable next step to analyse the two indexing approaches ina high minutiae quality environment.

Full pattern fingerprint SMR Inspired by the results of the full vascular pattern SMR, thisconcept could also be applied to the fingerprint modality. [XVK+08] hinted at performanceissues with poor quality fingerprints. Employing more of the fingerprint pattern in the

4In the means of low minutiae count with many spurious minutiae.

128

SMR could overcome this problem and enable the usage of this representation for poorquality data.

10.4 SummaryIn this chapter, the results of the project have been discussed in detail. First, the poor per-formance on the CASIA and PUT was discussed, followed by a discussion of several observationsalong the project. After a comparison of the experiments with state-of-the-art approaches, oneissue with the workload reporting was outlined.

Using the SMR for palm vein (and general vascular pattern) biometric systems offers amultitude of future research possibilities, several of which have been briefly outlined in thischapter.

129

Chapter 11

Conclusions

The interest in biometrics among governmental and industrial operators of access controlsystems has been steadily growing in recent years. With the adoption of biometrics in con-sumer products like notebooks or smartphones, the wide public has accepted biometrics andstrengthened their trust in such systems. Different huge biometric systems are deployed world-wide nowadays. Said systems are required to deliver high biometric performance while keepinga low computational workload profile. These trends make not only biometrics a relevant andattractive research area, but also workload reduction to increase the efficiency of high biomet-ric performance approaches. This thesis has pertained to the topic of workload reduction forbiometric identification in large-scale palm vein databases.

A recently-published biometric indexing approach based on Bloom filters and binary searchtrees for large-scale iris databases was adopted for the practical work for this thesis. To adoptthis indexing approach, the vascular pattern of the raw palm vein images was extracted usingan proposed signal processing subsystem. The minutiae - the endpoints and bifurcations - ofthe extracted vascular pattern where then transformed using a Fourier transformation basedapproach originally presented for the fingerprint characteristic. When transforming the real-valued representation yielded by the Fourier transformation to a binary form, it is possible toapply the Bloom filter-indexing. After implementing the Bloom filter system with careful con-figuration of the components, it has been demonstrated that the system is capable of achieving abiometric performance close to the baseline achieved with a naïve implementation of the Fourierrepresentation, while reducing the necessary workload by an additional ∼60% compared to anaïve implementation using the reduced binary Fourier representation. Some of the approachesused by the Bloom filter were not feasible and the fuzziness of the vascular pattern prevented ahigher workload reduction without losing too much biometric performance. However, the mostimportant approaches have been successfully applied, thus the system appears to be scalablein terms of workload reduction, biometric performance and enrollees.

An additional, less complex, biometric indexing approach merely using a reduced form of

130

the binary Fourier transformation representation and binary search trees has been presented. Itadopts most workload reduction strategies that are used for the Bloom filter-indexing approach.In three of four biometric performance metrics, it outperformed the Bloom filter-indexing whilestill reducing the workload compared to a naïve implementation using the reduced binaryFourier representation by ∼40%. Since the presented approach follows the same theory andimplementations as the binary search trees of the Bloom filter-indexing, it also appears to bescalable in terms of workload reduction, biometric performance and enrollees.

The overall workload is reduced to an average of 0.5% compared to the baseline of the naïveimplementation using the Fourier representation in both systems.

Facing the acceptable but not optimal result in the means of biometric performance yieldedby the baseline Fourier representation, several further research topics were presented to addressthis issue. The most promising topic was briefly tested and already achieved a much betterbiometric performance without any optimisation or tuning. Therefore, the system is not suitableyet for real-world palm vein system deployments but appears to be a promising base for furtherresearch with the proposed topics. Furthermore, the workload reduction achieved very promisingresults, which were limited by the performance of the base system. Since the base systemachieved a very high biometric performance for fingerprints, the workload reduction approachescan be adopted to the fingerprint modalities and should yield even better results in terms ofworkload reduction.

131

Appendix A

Appendix

A.1 Bloom filter-indexing - Quick traversal direction de-cision first child favouring

Observe line 7 in algorithm 6: the next decision will be determined with the next-to-last scorewhen the second comparison is skipped. This falsifies the decision making process because it isto be excepted, that the score of the comparison will be higher than the next-to-last score. Inother words, it is very likely, that the decision will always favour the first child if the previousdecision selected the second child.

Algorithm 6 Tree traversal for a single Bloom filter tree with single comparison strategy.1: function TraverseTree(Node, Query, Threshold, LastScore)2: if Node[left] = nil and Node[right] = nil then3: scoreLeft← similarity(Node[left], Query)4: if scoreLeft > LastScore then5: return TraverseTree(Node[left], Query, Threshold, scoreLeft)6: else7: return TraverseTree(Node[right], Query, Threshold, LastScore)8: end if9: end if10: score← similarity(Node[this], Query)11: if score > Threshold and score > LastScore then ▷ Second condition optional12: return Node[this]13: end if14: return nil15: end function

A.2 Experiments

132

Figure A.1: Used ROI for the PUT dataset.

10−2 10−1 1000.6

0.7

0.8

0.9

1

TP 0.1

False Positive Identification Rate

True

Posit

ive

Iden

tifica

tion

Rat

e

MC ×1MC ×2MC ×3MC ×4MC ×5MC ×6MC ×7MC ×8

Figure A.2: QSML ROC for the PolyU dataset with different maximum-curvature output gains (mul-tiplication).

133

10−3 10−2 10−1 100

0.1

0.2

0.3

0.4

0.50.5

0.6

0.7

0.8

0.9

1

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

SMC (Absolute)SMC (Real)SML (Absolute)SML (Real)QSMC (Absolute)QSMC (Real)QSML (Absolute)QSML (Real)

Figure A.3: ROC curves of the SML, SMC, QSML and QSMC type for the first session of the PolyUdataset.

SMR Ident. Verif.

Type Pre-Selection λmin λmax TP0.5 TP0.1 TP0 EER

SML qM*> 0.2 0.1 0.51 91.2% 91.1% 90.4% 2.1%

QSML* 0.1 0.55 91.4% 91.1% 90.3% 2.1%SML 0.1 0.60 90.9% 90.5% 90.0% 2.2%* Maximum curvature ×7.

Table A.1: Top three configurations (TP0.1), for a biometric identification system using the first sessionPolyU dataset, with their corresponding verification EER, ordered by TP0.

L TP0.5 TP0.1 TP0 EER

5 91.3% 90.9% 89.7% 2.2%15 91.3% 90.6% 90.6% 2.1%25 91.6% 90.9% 90.3% 2.1%50 91.8% 90.6% 90.4% 2.1%100 91.4% 90.7% 90.2% 2.1%

Table A.2: TP0, TP0.1, TP0.5 and EER for the different CPCA feature reduction training set sizes(L).

134

10−3 10−2 10−1 1000.50.5

0.6

0.7

0.8

0.9

1

5 · 10−3

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhoodqM > 0.1qM > 0.2Bifurc. & qM > 0.1

(a) SML

10−3 10−2 10−1 1000.50.5

0.6

0.7

0.8

0.9

1

5 · 10−3

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhood

(b) QSML

10−3 10−2 10−1 1000.50.5

0.6

0.7

0.8

0.9

1

5 · 10−3

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhoodqM > 0.1qM > 0.2Bifurc. & qM > 0.1Bifurc. & qM > 0.2

(c) SMC

10−3 10−2 10−1 1000.2

0.3

0.4

0.50.5

0.6

0.7

0.8

0.9

1

5 · 10−3

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate

All MinutiaeOnly BifurcationsNeighbourhood

(d) QSMC

Figure A.4: ROC curves of the real-valued SML, SMC, QSML and QSMC type for the first session ofthe PolyU dataset with applied minutiae selection.

10−3 10−2 10−1 100

0.9

0.91

0.92

0.93

0.94

0.95

0.960.96

5 · 10−3

False Positive Identification Rate

Tru

eP

osit

ive

Iden

tific

atio

nR

ate QSML Tuned

QSMLSML (qM > 0.2) TunedSML (qM > 0.2)SML (Tuned)

Figure A.5: ROC curves of the most promising SMR types, tuned and untuned, for the first sessionof the PolyU dataset.

135

Bit-error Bloom filter SMR Workload Recognition Performance

correction W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

Majority Voting

4 4 0.1 1.31 ∗ 106 0.4883% 100% 70.2% 65.4% 61.1% 29.3%6 4 0.2 8.73 ∗ 105 0.3255% 66.67% 66.5% 61.5% 59.5% 30.9%4 4 0.3 1.31 ∗ 106 0.4883% 100% 63.2% 61.9% 57.5% 32.9%8 4 0.4 6.55 ∗ 105 0.2441% 50% 57.3% 53.4% 53% 37.4%5 4 0.5 1.04 ∗ 106 0.3906% 80% 53.1% 48.4% 45.7% 44.7%8 5 0.6 1.04 ∗ 106 0.3906% 80% 49.8% 42.3% 34.8% 55.6%6 5 0.7 1.39 ∗ 106 0.5208% 106.7% 45.2% 42% 39.6% 50.8%8 4 0.8 6.55 ∗ 105 0.2441% 50% 46.9% 34.1% 32.2% 58.2%6 5 0.9 1.39 ∗ 106 0.5208% 106.7% 33.6% 28.4% 26.9% 63.5%

Transposed

4 5 0.1 2.09 ∗ 106 0.7812% 160% 14.9% 13.2% 11.8% 78.6%4 4 0.2 1.31 ∗ 106 0.4883% 100% 13.4% 11.4% 8.75% 81.7%4 5 0.3 2.09 ∗ 106 0.7812% 160% 14.5% 13.4% 10.4% 80%4 4 0.4 1.31 ∗ 106 0.4883% 100% 15.9% 9.92% 6.72% 83.7%4 4 0.5 1.31 ∗ 106 0.4883% 100% 15.6% 11% 5.47% 84.9%4 4 0.6 1.31 ∗ 106 0.4883% 100% 11.5% 7.97% 6.88% 83.5%4 4 0.7 1.31 ∗ 106 0.4883% 100% 9.61% 7.11% 5.47% 85.9%4 4 0.8 1.31 ∗ 106 0.4883% 100% 7.58% 6.09% 5% 85.4%4 4 0.9 1.31 ∗ 106 0.4883% 100% 6.72% 5.39% 3.75% 86.7%

Transposed&

Majority Voting

4 5 0.1 2.09 ∗ 106 0.7812% 160% 17% 14% 10.1% 80.3%4 4 0.2 1.31 ∗ 106 0.4883% 100% 14.5% 9.61% 9.61% 80.8%4 5 0.3 2.09 ∗ 106 0.7812% 160% 15.2% 13.2% 12.1% 78.3%4 4 0.4 1.31 ∗ 106 0.4883% 100% 17% 10.3% 7.42% 83%4 4 0.5 1.31 ∗ 106 0.4883% 100% 16.2% 9.53% 7.11% 83.3%4 4 0.6 1.31 ∗ 106 0.4883% 100% 11.5% 7.66% 7.66% 82.7%4 4 0.7 1.31 ∗ 106 0.4883% 100% 9.06% 6.48% 5.47% 84.9%4 4 0.8 1.31 ∗ 106 0.4883% 100% 7.58% 5.7% 4.45% 85.9%4 4 0.9 1.31 ∗ 106 0.4883% 100% 6.02% 5.16% 4.38% 86%

* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table A.3: Best achieved TP0 with corresponding TP0.1 and TP0.5 for every tested binarisation maskthreshold with their corresponding H and W (T = 64) using the binary PSML-CPCA with severalbit-error reduction strategies, ordered by mask threshold.

136

Bloom filter SMR Workload Recognition Performance

T W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

16

6 4 0.1 4.42 ∗ 105 0.1647% 33.72% 82.3% 82% 80.9% 9.5%4 4 0.2 6.60 ∗ 105 0.246% 50.39% 84.5% 83.8% 83% 7.43%5 4 0.3 5.29 ∗ 105 0.1972% 40.39% 83.2% 82.6% 82.3% 8.06%4 6 0.4 1.75 ∗ 106 0.6529% 133.7% 83% 82.7% 82.5% 7.9%4 6 0.5 1.75 ∗ 106 0.6529% 133.7% 83% 82.5% 82.3% 8.06%5 5 0.6 8.43 ∗ 105 0.3144% 64.39% 83.4% 82.5% 82.2% 8.21%5 4 0.7 5.29 ∗ 105 0.1972% 40.39% 80.8% 80.5% 79.8% 10.6%5 5 0.8 8.43 ∗ 105 0.3144% 64.39% 82.8% 82.6% 81% 9.4%6 5 0.9 7.04 ∗ 105 0.2623% 53.72% 80.5% 79.1% 78.8% 11.6%

32

6 4 0.1 6.60 ∗ 105 0.246% 50.39% 87% 86.4% 85.1% 5.32%4 7 0.2 4.49 ∗ 106 1.676% 343.2% 84.2% 83.4% 83% 7.35%6 6 0.3 1.75 ∗ 106 0.6529% 133.7% 85.9% 85.2% 84.1% 6.26%8 5 0.4 7.91 ∗ 105 0.2949% 60.39% 87% 86.6% 86.1% 4.31%7 4 0.5 5.66 ∗ 105 0.2112% 43.25% 87.1% 86.1% 85.7% 4.7%6 6 0.6 1.75 ∗ 106 0.6529% 133.7% 85.8% 85.2% 83.9% 6.49%5 4 0.7 7.91 ∗ 105 0.2949% 60.39% 87.8% 87.6% 86.6% 3.84%5 5 0.8 1.26 ∗ 106 0.4707% 96.39% 87.7% 87.3% 86.2% 4.15%5 4 0.9 7.91 ∗ 105 0.2949% 60.39% 87.1% 85.6% 84.8% 5.63%

* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.‡ Full table appended in A.4.

Table A.4: Achieved PSML-CPCA Bloom filter TP0 with corresponding TP0.1 and TP0.5, with a finalPSML-CPCA comparison using T = 32 and T = 16 trees.

137

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 4 4 0.2 3.53 ∗ 105 0.1316% 26.9% 78.2% 77.3% 77.1% 13.30%2 4 4 0.3 3.73 ∗ 105 0.1392% 28.5% 82.0% 81.4% 81.3% 9.10%3 4 4 0.4 3.94 ∗ 105 0.1469% 30.1% 84.7% 84.3% 83.4% 6.96%4 4 4 0.4 4.14 ∗ 105 0.1545% 31.6% 85.8% 85.5% 84.5% 5.95%5 4 4 0.2 4.35 ∗ 105 0.1621% 33.2% 86.2% 85.3% 84.6% 5.79%6 5 5 0.6 5.81 ∗ 105 0.2168% 44.4% 87.3% 86.6% 85.3% 5.09%7 5 5 0.6 6.08 ∗ 105 0.2265% 46.4% 88.0% 87.2% 85.9% 4.54%8 5 5 0.6 6.34 ∗ 105 0.2363% 48.4% 88.4% 87.4% 86.1% 4.31%9 5 5 0.6 6.60 ∗ 105 0.2460% 50.4% 88.4% 87.4% 86.2% 4.15%10 4 4 0.2 5.37 ∗ 105 0.2003% 41.0% 88.0% 86.6% 86.2% 4.15%11 4 4 0.2 5.58 ∗ 105 0.2079% 42.6% 88.0% 86.6% 86.2% 4.15%12 5 5 0.6 7.39 ∗ 105 0.2753% 56.4% 88.6% 88.0% 86.5% 3.92%13 5 5 0.6 7.65 ∗ 105 0.2851% 58.4% 88.6% 88.0% 86.5% 3.92%14 5 5 0.6 7.91 ∗ 105 0.2949% 60.4% 88.8% 88.1% 86.6% 3.76%15 5 5 0.6 8.17 ∗ 105 0.3046% 62.4% 89.1% 88.4% 87.5% 2.90%16 5 5 0.6 8.43 ∗ 105 0.3144% 64.4% 89.0% 88.4% 87.4% 2.98%17 5 5 0.6 8.70 ∗ 105 0.3242% 66.4% 89.1% 88.4% 87.5% 2.90%18 5 5 0.6 8.96 ∗ 105 0.3339% 68.4% 89.2% 88.6% 87.7% 2.74%19 5 5 0.6 9.22 ∗ 105 0.3437% 70.4% 89.5% 88.9% 88.0% 2.43%20 5 5 0.6 9.48 ∗ 105 0.3535% 72.4% 89.5% 88.9% 88.0% 2.43%21 5 5 0.6 9.75 ∗ 105 0.3632% 74.4% 89.5% 88.9% 88.0% 2.43%22 7 4 0.5 4.49 ∗ 105 0.1676% 34.3% 87.9% 87.0% 86.7% 3.68%23 5 6 0.4 1.70 ∗ 106 0.6367% 130.4% 87.4% 87.2% 86.2% 4.23%24 5 6 0.4 1.75 ∗ 106 0.6529% 133.7% 87.3% 87.1% 86.2% 4.23%25 5 6 0.4 1.79 ∗ 106 0.6692% 137.1% 87.3% 87.0% 86.1% 4.31%26 5 6 0.4 1.84 ∗ 106 0.6855% 140.4% 87.3% 87.0% 86.1% 4.31%27 5 6 0.4 1.88 ∗ 106 0.7018% 143.7% 87.3% 87.1% 86.2% 4.23%28 5 6 0.4 1.92 ∗ 106 0.7181% 147.1% 87.3% 87.1% 86.2% 4.23%29 5 6 0.4 1.97 ∗ 106 0.7343% 150.4% 87.3% 87.1% 86.2% 4.23%30 5 6 0.4 2.01 ∗ 106 0.7506% 153.7% 87.3% 87.1% 86.2% 4.23%31 5 6 0.4 2.05 ∗ 106 0.7669% 157.1% 87.3% 87.1% 86.2% 4.23%32 5 6 0.4 2.10 ∗ 106 0.7832% 160.4% 87.3% 87.1% 86.2% 4.23%33 5 6 0.4 2.14 ∗ 106 0.7994% 163.7% 87.3% 87.1% 86.2% 4.23%34 5 6 0.4 2.18 ∗ 106 0.8157% 167.1% 87.3% 87.1% 86.2% 4.23%35 5 6 0.4 2.23 ∗ 106 0.8320% 170.4% 87.3% 87.1% 86.2% 4.23%36 5 6 0.4 2.27 ∗ 106 0.8483% 173.7% 87.3% 87.1% 86.2% 4.23%37 5 6 0.4 2.32 ∗ 106 0.8645% 177.1% 87.3% 87.1% 86.2% 4.23%38 5 6 0.4 2.36 ∗ 106 0.8808% 180.4% 87.3% 87.1% 86.2% 4.23%39 5 6 0.4 2.40 ∗ 106 0.8971% 183.7% 87.3% 87.1% 86.2% 4.23%40 5 6 0.4 2.45 ∗ 106 0.9134% 187.1% 87.3% 87.1% 86.2% 4.23%41 5 6 0.4 2.49 ∗ 106 0.9296% 190.4% 87.3% 87.1% 86.2% 4.23%42 5 6 0.4 2.53 ∗ 106 0.9459% 193.7% 87.3% 87.1% 86.2% 4.23%43 5 6 0.4 2.58 ∗ 106 0.9622% 197.1% 87.3% 87.1% 86.2% 4.23%44 5 6 0.4 2.62 ∗ 106 0.9785% 200.4% 87.3% 87.1% 86.2% 4.23%45 5 6 0.4 2.67 ∗ 106 0.9947% 203.7% 87.3% 87.1% 86.2% 4.23%46 5 6 0.4 2.71 ∗ 106 1.0110% 207.1% 87.3% 87.1% 86.2% 4.23%47 5 6 0.4 2.75 ∗ 106 1.0270% 210.4% 87.3% 87.1% 86.2% 4.23%48 5 6 0.4 2.80 ∗ 106 1.0440% 213.7% 87.3% 87.1% 86.2% 4.23%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table A.5: Best achieved PSML-CPCA Bloom filter TP0 with corresponding TP0.1 and TP0.5, witha final binary PSML-CPCA comparison and tree pre-selection of t = 1, . . . , 48 trees at T = 64.

138

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 4 5 0.1 3.16 ∗ 105 0.1179% 24.14% 66.6% 65.9% 65.7% 24.7%2 4 5 0.1 3.65 ∗ 105 0.1362% 27.89% 74.2% 73.8% 73.2% 17.2%3 4 5 0.3 4.14 ∗ 105 0.1545% 31.64% 76.6% 76.2% 76.0% 14.4%4 4 5 0.4 4.63 ∗ 105 0.1728% 35.39% 79.3% 79.1% 78.4% 12%5 4 5 0.5 5.13 ∗ 105 0.1911% 39.14% 80.7% 80.2% 80.2% 10.2%6 4 4 0.3 3.53 ∗ 105 0.1316% 26.95% 82.3% 81.8% 81.6% 8.76%7 4 4 0.3 3.84 ∗ 105 0.1431% 29.30% 83.5% 83.2% 83.1% 7.28%8 4 5 0.5 6.60 ∗ 105 0.2460% 50.39% 84.3% 83.8% 83.8% 6.65%9 4 5 0.5 7.09 ∗ 105 0.2644% 54.14% 85.1% 84.5% 84.5% 5.95%10 4 5 0.5 7.58 ∗ 105 0.2827% 57.89% 85.5% 84.8% 84.8% 5.63%11 4 5 0.5 8.07 ∗ 105 0.3010% 61.64% 85.9% 85.1% 85.1% 5.32%12 4 5 0.5 8.57 ∗ 105 0.3193% 65.39% 86.3% 85.5% 85.5% 4.85%13 4 5 0.6 9.06 ∗ 105 0.3376% 69.14% 87.3% 86.4% 86.1% 4.31%14 4 5 0.6 9.55 ∗ 105 0.3559% 72.89% 87.5% 86.6% 86.2% 4.15%15 4 5 0.6 1.00 ∗ 106 0.3742% 76.64% 87.8% 86.9% 86.4% 3.99%16 4 4 0.5 6.60 ∗ 105 0.2460% 50.39% 87.7% 86.7% 86.6% 3.76%17 4 5 0.6 1.10 ∗ 106 0.4108% 84.14% 88.1% 87.2% 86.7% 3.68%18 4 5 0.6 1.15 ∗ 106 0.4292% 87.89% 88.1% 87.2% 86.7% 3.68%19 4 5 0.6 1.20 ∗ 106 0.4475% 91.64% 88.0% 87.3% 86.1% 4.31%20 4 4 0.5 7.83 ∗ 105 0.2918% 59.77% 88.0% 87.1% 86.1% 4.31%21 4 4 0.5 8.14 ∗ 105 0.3033% 62.11% 88.0% 87.1% 86.1% 4.31%22 4 4 0.5 8.44 ∗ 105 0.3147% 64.45% 88.0% 87.2% 86.2% 4.15%23 6 5 0.6 9.33 ∗ 105 0.3478% 71.22% 87.7% 86.6% 86.6% 3.84%24 4 4 0.5 9.06 ∗ 105 0.3376% 69.14% 88.2% 87.3% 86.4% 3.99%25 4 4 0.5 9.36 ∗ 105 0.3490% 71.48% 88.2% 87.3% 86.4% 3.99%26 5 4 0.7 7.75 ∗ 105 0.2888% 59.14% 87.4% 87.2% 86.2% 4.15%27 5 4 0.7 7.99 ∗ 105 0.2979% 61.02% 87.5% 87.3% 86.3% 4.07%28 5 4 0.7 8.24 ∗ 105 0.3071% 62.89% 87.6% 87.3% 86.4% 3.99%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table A.6: Best achieved PSML-CPCA Bloom filter TP0 with corresponding TP0.1 and TP0.5, witha final binary PSML-CPCA comparison and tree pre-selection of t = 1, . . . , 28 trees at T = 32.

139

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 4 4 0.2 3.53 ∗ 105 0.1316% 26.95% 78.9% 78.4% 78.0% 12.4%2 4 5 0.3 5.94 ∗ 105 0.2216% 45.39% 83.6% 83.1% 82.7% 7.67%3 4 4 0.4 3.94 ∗ 105 0.1469% 30.08% 85.4% 84.7% 84.6% 5.79%4 4 4 0.4 4.14 ∗ 105 0.1545% 31.64% 87.0% 86.2% 86.1% 4.31%5 4 4 0.5 4.35 ∗ 105 0.1621% 33.20% 87.6% 87.3% 86.6% 3.84%6 4 4 0.5 4.55 ∗ 105 0.1698% 34.77% 88.2% 87.9% 87.2% 3.21%7 4 4 0.5 4.76 ∗ 105 0.1774% 36.33% 88.8% 88.4% 87.7% 2.67%8 4 4 0.5 4.96 ∗ 105 0.1850% 37.89% 89.0% 88.7% 88.0% 2.35%9 4 4 0.5 5.17 ∗ 105 0.1926% 39.45% 89.3% 89.0% 88.4% 2.04%10 4 4 0.5 5.37 ∗ 105 0.2003% 41.02% 89.6% 89.3% 88.4% 1.96%11 4 4 0.5 5.58 ∗ 105 0.2079% 42.58% 89.9% 89.6% 88.7% 1.73%12 4 4 0.5 5.78 ∗ 105 0.2155% 44.14% 90.1% 89.8% 88.8% 1.57%13 4 4 0.5 5.99 ∗ 105 0.2232% 45.70% 89.9% 88.9% 88.8% 1.57%14 5 4 0.5 4.96 ∗ 105 0.1850% 37.89% 89.5% 88.7% 88.6% 1.81%15 5 4 0.5 5.13 ∗ 105 0.1911% 39.14% 89.5% 88.7% 88.6% 1.81%16 5 4 0.5 5.29 ∗ 105 0.1972% 40.39% 89.5% 88.7% 88.6% 1.81%17 5 4 0.5 5.45 ∗ 105 0.2033% 41.64% 89.5% 88.8% 88.7% 1.73%18 4 5 0.6 1.11 ∗ 106 0.4169% 85.39% 90.2% 89.4% 88.8% 1.65%19 4 5 0.6 1.15 ∗ 106 0.4292% 87.89% 90.2% 89.4% 88.8% 1.65%20 4 5 0.6 1.18 ∗ 106 0.4414% 90.39% 90.2% 89.4% 88.8% 1.65%21 4 5 0.6 1.21 ∗ 106 0.4536% 92.89% 90.3% 89.5% 88.8% 1.57%22 6 4 0.5 5.23 ∗ 105 0.1952% 39.97% 89.9% 89.1% 89.0% 1.42%23 6 4 0.5 5.37 ∗ 105 0.2003% 41.02% 89.8% 89.0% 88.9% 1.49%24 6 4 0.5 5.51 ∗ 105 0.2054% 42.06% 89.8% 89.0% 88.9% 1.49%25 6 4 0.5 5.64 ∗ 105 0.2104% 43.10% 89.8% 89.1% 89.0% 1.42%26 6 4 0.5 5.78 ∗ 105 0.2155% 44.14% 89.9% 89.1% 89.0% 1.42%27 6 4 0.5 5.92 ∗ 105 0.2206% 45.18% 90.0% 89.1% 89.0% 1.42%28 4 5 0.6 1.44 ∗ 106 0.5390% 110.4% 90.2% 89.3% 88.7% 1.73%29 4 5 0.6 1.47 ∗ 106 0.5512% 112.9% 90.2% 89.3% 88.7% 1.73%30 4 5 0.6 1.51 ∗ 106 0.5634% 115.4% 90.2% 89.3% 88.7% 1.73%31 4 5 0.6 1.54 ∗ 106 0.5756% 117.9% 89.9% 89.3% 88.7% 1.73%32 4 5 0.6 1.57 ∗ 106 0.5878% 120.4% 89.9% 89.3% 88.7% 1.73%33 4 5 0.6 1.61 ∗ 106 0.6001% 122.9% 89.9% 89.3% 88.7% 1.73%34 4 5 0.6 1.64 ∗ 106 0.6123% 125.4% 89.9% 89.3% 88.7% 1.73%35 4 5 0.6 1.67 ∗ 106 0.6245% 127.9% 90.0% 89.4% 88.8% 1.65%36 4 5 0.6 1.70 ∗ 106 0.6367% 130.4% 90.0% 89.4% 88.8% 1.65%37 4 5 0.6 1.74 ∗ 106 0.6489% 132.9% 90.0% 89.4% 88.8% 1.65%38 4 5 0.6 1.77 ∗ 106 0.6611% 135.4% 90.0% 89.4% 88.8% 1.65%39 4 5 0.6 1.80 ∗ 106 0.6733% 137.9% 90.0% 89.4% 88.8% 1.65%40 4 5 0.6 1.84 ∗ 106 0.6855% 140.4% 90.0% 89.4% 88.8% 1.65%41 4 5 0.6 1.87 ∗ 106 0.6977% 142.9% 90.0% 89.4% 88.8% 1.65%42 4 5 0.6 1.90 ∗ 106 0.7099% 145.4% 90.0% 89.4% 88.8% 1.65%43 6 5 0.6 1.29 ∗ 106 0.4821% 98.72% 90.2% 89.4% 88.8% 1.65%44 6 5 0.6 1.31 ∗ 106 0.4902% 100.4% 90.2% 89.4% 88.8% 1.65%45 6 5 0.6 1.33 ∗ 106 0.4983% 102.1% 90.2% 89.4% 88.8% 1.65%46 6 5 0.6 1.35 ∗ 106 0.5065% 103.7% 90.2% 89.4% 88.8% 1.65%47 4 5 0.6 2.06 ∗ 106 0.7710% 157.9% 90.0% 89.4% 88.8% 1.65%48 4 5 0.6 2.10 ∗ 106 0.7832% 160.4% 90.0% 89.4% 88.8% 1.65%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table A.7: Best achieved binary PSML-CPCA Bloom filter TP0 with corresponding TP0.1 and TP0.5,with a final real-valued PSML-CPCA comparison and tree pre-selection of t = 1, . . . , 48 trees atT = 64.

140

Bit-error Bloom filter SMR Workload Recognition Performance

correction W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

Transposed

4 5 0.1 1.34 ∗ 107 5.0% 160% 84.5% 84.1% 83% 7.43%4 4 0.2 8.38 ∗ 106 3.125% 100% 84.4% 82.8% 82.1% 8.29%6 4 0.3 5.59 ∗ 106 2.083% 66.67% 83.8% 81.9% 81.6% 8.76%6 4 0.4 5.59 ∗ 106 2.083% 66.67% 83.5% 81.4% 80.5% 9.9%7 6 0.5 1.27 ∗ 107 4.762% 152.4% 79.2% 78.4% 77.6% 12.8%7 5 0.6 7.66 ∗ 106 2.857% 91.43% 80.9% 77.7% 76.7% 13.7%6 4 0.7 5.59 ∗ 106 2.083% 66.67% 77.1% 74.6% 73% 17.4%8 4 0.8 4.19 ∗ 106 1.562% 50% 76.8% 73.4% 67% 23.4%8 6 0.9 1.11 ∗ 107 4.167% 133.3% 68% 65% 60.2% 30.2%

Majority Voting

5 4 0.1 6.71 ∗ 106 2.5% 80% 84.7% 83% 81.6% 8.76%5 4 0.2 6.71 ∗ 106 2.5% 80% 84.7% 84% 81.6% 8.76%6 4 0.3 5.59 ∗ 106 2.083% 66.67% 84.1% 82% 80.9% 9.5%7 4 0.4 4.79 ∗ 106 1.786% 57.14% 81.5% 80.9% 79.8% 10.6%7 6 0.5 1.27 ∗ 107 4.762% 152.4% 81.2% 79.3% 78.4% 12%7 4 0.6 4.79 ∗ 106 1.786% 57.14% 81.7% 77.5% 76.1% 14.3%7 4 0.7 4.79 ∗ 106 1.786% 57.14% 78.7% 74.5% 72.7% 17.7%8 8 0.8 3.35 ∗ 107 12.5% 400% 72.4% 68.9% 66.1% 24.3%8 8 0.9 3.35 ∗ 107 12.5% 400% 68.4% 62% 54.5% 35.9%

Transposed&

Majority Voting

4 5 0.1 1.34 ∗ 107 5.0% 160% 84.3% 83.4% 82.9% 7.51%4 4 0.2 8.38 ∗ 106 3.125% 100% 84.8% 82.9% 82.3% 8.13%6 6 0.3 1.49 ∗ 107 5.556% 177.8% 82.3% 81.3% 79.7% 10.7%6 7 0.4 2.55 ∗ 107 9.524% 304.8% 83.2% 81.2% 79.4% 11%7 6 0.5 1.27 ∗ 107 4.762% 152.4% 79.5% 78.2% 77.5% 12.9%7 5 0.6 7.66 ∗ 106 2.857% 91.43% 80.2% 77.7% 76.4% 14%6 4 0.7 5.59 ∗ 106 2.083% 66.67% 76.9% 75.9% 71.6% 18.8%8 4 0.8 4.19 ∗ 106 1.562% 50% 77.2% 73.2% 66.5% 23.9%7 4 0.9 4.79 ∗ 106 1.786% 57.14% 70.3% 67.8% 61.9% 28.5%

* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table A.8: Top three achieved TP0 with corresponding TP0.1 and TP0.5 for every tested binarisationmask threshold with their corresponding H and W (T = 64) using the binary PSML with severalbit-error reduction strategies, ordered by mask threshold.

141

Bit-error Bloom filter SMR Workload Recognition Performance

correction W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

None

8 4 0.1 4.22 ∗ 106 0.7874% 50.39% 87.8% 86.9% 86.5% 3.92%8 6 0.2 1.12 ∗ 107 2.089% 133.7% 88.7% 87.5% 87.1% 3.29%5 5 0.3 1.07 ∗ 107 2.006% 128.4% 89.5% 88.8% 86.8% 3.6%7 4 0.4 4.82 ∗ 106 0.899% 57.53% 89.8% 89% 87.4% 2.98%9 4 0.5 3.76 ∗ 106 0.7005% 44.84% 89.2% 88.8% 88.8% 1.65%8 6 0.6 1.12 ∗ 107 2.089% 133.7% 89.8% 89.5% 89% 1.42%7 7 0.7 2.19 ∗ 107 4.088% 261.6% 90.1% 89.1% 88.9% 1.49%7 5 0.8 7.70 ∗ 106 1.435% 91.82% 89.2% 88.8% 88.6% 1.81%8 4 0.9 4.22 ∗ 106 0.7874% 50.39% 89.8% 88.7% 87.8% 2.59%

Transposed

8 5 0.1 6.74 ∗ 106 1.256% 80.39% 88% 87.5% 86.6% 3.76%5 5 0.2 1.07 ∗ 107 2.006% 128.4% 88.5% 87.7% 87.3% 3.13%8 12 0.3 3.57 ∗ 108 66.67% 4267% 88% 87.6% 87.3% 3.06%7 4 0.4 4.82 ∗ 106 0.899% 57.53% 89.2% 87.7% 87.5% 2.9%6 6 0.5 1.49 ∗ 107 2.784% 178.2% 89.4% 88.9% 88.5% 1.88%5 5 0.6 1.07 ∗ 107 2.006% 128.4% 89.8% 88.7% 88% 2.35%5 8 0.7 5.37 ∗ 107 10.01% 640.4% 89.8% 88.8% 87.9% 2.51%5 8 0.8 5.37 ∗ 107 10.01% 640.4% 89.4% 88.3% 87.5% 2.9%5 8 0.9 5.37 ∗ 107 10.01% 640.4% 89% 87.6% 86.9% 3.53%

MajorityVoting

8 4 0.1 4.22 ∗ 106 0.7874% 50.39% 87.7% 86.8% 86.4% 3.99%7 7 0.2 2.19 ∗ 107 4.088% 261.6% 88.5% 87.4% 87% 3.37%5 5 0.3 1.07 ∗ 107 2.006% 128.4% 89.5% 88.8% 86.8% 3.6%5 5 0.4 1.07 ∗ 107 2.006% 128.4% 89.8% 89% 87.4% 2.98%5 5 0.5 1.07 ∗ 107 2.006% 128.4% 89.3% 89% 88.6% 1.81%6 5 0.6 8.98 ∗ 106 1.673% 107.1% 90.1% 89.1% 88.5% 1.88%7 4 0.7 4.82 ∗ 106 0.899% 57.53% 90.2% 89.4% 88.6% 1.81%8 5 0.8 6.74 ∗ 106 1.256% 80.39% 89.8% 89.5% 88.9% 1.49%8 5 0.9 6.74 ∗ 106 1.256% 80.39% 89.3% 88.2% 87.7% 2.74%

* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table A.9: Top three achieved TP0 with corresponding TP0.1 and TP0.5 using the binary PSML withand without bit-error reduction strategies, ordered by mask threshold.

142

Bloom filter SMR Workload Recognition Performance

T W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

16

4 4 0.1 4.22 ∗ 106 0.7874% 50.39% 80.5% 79.6% 79.5% 10.9%4 5 0.2 6.74 ∗ 106 1.256% 80.39% 81.2% 80.6% 78.6% 11.8%5 5 0.3 5.40 ∗ 106 1.006% 64.39% 81.3% 80.9% 79.7% 10.7%6 5 0.4 4.50 ∗ 106 0.8394% 53.72% 83.1% 83% 81.1% 9.3%5 5 0.5 5.40 ∗ 106 1.006% 64.39% 82.4% 81.8% 81.6% 8.84%4 5 0.6 6.74 ∗ 106 1.256% 80.39% 84.1% 83.8% 83.3% 7.12%4 6 0.7 1.12 ∗ 107 2.089% 133.7% 83.3% 82.5% 81.9% 8.53%5 5 0.8 5.40 ∗ 106 1.006% 64.39% 82.5% 82.3% 80.9% 9.5%5 4 0.9 3.38 ∗ 106 0.6311% 40.39% 81.8% 81.3% 80.8% 9.6%

32

4 4 0.1 6.32 ∗ 106 1.178% 75.39% 86.8% 86.2% 85.9% 4.54%4 9 0.2 8.95 ∗ 107 16.67% 1067% 87.2% 86.8% 85.9% 4.46%4 8 0.3 5.03 ∗ 107 9.381% 600.4% 88.2% 87.1% 86.1% 4.31%6 5 0.4 6.74 ∗ 106 1.256% 80.39% 89.1% 88.2% 86.9% 3.53%7 4 0.5 3.62 ∗ 106 0.6757% 43.25% 89.2% 88.4% 87.9% 2.51%5 4 0.6 5.06 ∗ 106 0.9436% 60.39% 89.1% 88.5% 87.8% 2.59%7 4 0.7 3.62 ∗ 106 0.6757% 43.25% 89% 88.1% 88% 2.35%8 5 0.8 5.06 ∗ 106 0.9436% 60.39% 88.4% 88% 87.6% 2.82%8 4 0.9 3.17 ∗ 106 0.592% 37.89% 89.2% 87.9% 87% 3.45%

* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML ω256.

Table A.10: Achieved PSML Bloom filter TP0 with corresponding TP0.1 and TP0.5, with a final binaryPSML comparison using T = 32 and T = 16 trees.

143

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 4 5 0.1 3.59 ∗ 106 0.6702% 42.89% 72.1% 72.0% 71.7% 18.7%2 4 5 0.2 3.80 ∗ 106 0.7092% 45.39% 79.1% 78.7% 78.0% 12.4%3 4 4 0.6 2.52 ∗ 106 0.4700% 30.08% 82.9% 82.1% 81.2% 9.2%4 4 5 0.6 4.22 ∗ 106 0.7874% 50.39% 84.7% 84.3% 83.6% 6.81%5 4 5 0.5 4.43 ∗ 106 0.8264% 52.89% 85.5% 84.9% 84.7% 5.71%6 4 5 0.6 4.64 ∗ 106 0.8655% 55.39% 86.8% 85.9% 85.3% 5.09%7 4 5 0.6 4.85 ∗ 106 0.9045% 57.89% 87.4% 86.6% 86.1% 4.31%8 6 4 0.8 2.12 ∗ 106 0.3967% 25.39% 87.6% 87.3% 87.1% 3.29%9 6 4 0.8 2.21 ∗ 106 0.4130% 26.43% 88.0% 87.7% 87.5% 2.9%10 6 4 0.8 2.30 ∗ 106 0.4293% 27.47% 88.3% 88.0% 87.7% 2.67%11 6 4 0.8 2.39 ∗ 106 0.4456% 28.52% 88.8% 88.4% 88.2% 2.2%12 4 5 0.6 5.90 ∗ 106 1.1000% 70.39% 89.0% 88.4% 87.7% 2.74%13 4 5 0.6 6.11 ∗ 106 1.1390% 72.89% 89.1% 88.6% 87.8% 2.59%14 5 5 0.6 5.06 ∗ 106 0.9436% 60.39% 89.1% 88.6% 87.9% 2.51%15 7 5 0.7 3.74 ∗ 106 0.6981% 44.68% 88.8% 88.2% 88.1% 2.28%16 7 5 0.7 3.86 ∗ 106 0.7204% 46.10% 89.1% 88.4% 88.4% 2.04%17 7 4 0.7 2.50 ∗ 106 0.4665% 29.85% 89.4% 88.6% 88.5% 1.88%18 7 4 0.7 2.57 ∗ 106 0.4804% 30.75% 89.5% 88.7% 88.6% 1.81%19 7 4 0.7 2.65 ∗ 106 0.4944% 31.64% 89.6% 88.8% 88.8% 1.65%20 7 4 0.7 2.72 ∗ 106 0.5083% 32.53% 89.8% 89.0% 88.9% 1.49%21 4 7 0.6 2.22 ∗ 107 4.1360% 264.7% 89.8% 89.4% 88.6% 1.81%22 5 7 0.6 1.82 ∗ 107 3.3990% 217.5% 89.9% 89.4% 88.6% 1.81%23 4 7 0.6 2.34 ∗ 107 4.3590% 279.0% 89.9% 89.5% 88.7% 1.73%24 5 7 0.6 1.92 ∗ 107 3.5780% 229.0% 90.0% 89.5% 88.7% 1.73%25 5 7 0.6 1.96 ∗ 107 3.6670% 234.7% 90.1% 89.5% 88.8% 1.65%26 5 7 0.6 2.01 ∗ 107 3.7560% 240.4% 90.1% 89.5% 88.8% 1.65%27 5 7 0.6 2.06 ∗ 107 3.8450% 246.1% 90.1% 89.5% 88.8% 1.65%28 5 7 0.6 2.11 ∗ 107 3.9350% 251.8% 90.1% 89.5% 88.8% 1.65%29 5 7 0.6 2.16 ∗ 107 4.0240% 257.5% 90.1% 89.5% 88.8% 1.65%30 5 7 0.6 2.20 ∗ 107 4.1130% 263.2% 90.1% 89.5% 88.8% 1.65%31 5 7 0.6 2.25 ∗ 107 4.2030% 269.0% 90.2% 89.6% 88.8% 1.57%32 5 7 0.6 2.30 ∗ 107 4.2920% 274.7% 90.2% 89.6% 88.8% 1.57%33 5 7 0.6 2.35 ∗ 107 4.3810% 280.4% 90.2% 89.6% 88.8% 1.57%34 5 7 0.6 2.40 ∗ 107 4.4700% 286.1% 90.2% 89.6% 88.8% 1.57%35 7 7 0.7 1.74 ∗ 107 3.2590% 208.6% 90.1% 89.1% 88.9% 1.49%36 7 7 0.7 1.78 ∗ 107 3.3220% 212.6% 90.1% 89.1% 88.9% 1.49%37 7 7 0.7 1.81 ∗ 107 3.3860% 216.7% 90.1% 89.1% 88.9% 1.49%38 7 7 0.7 1.85 ∗ 107 3.4500% 220.8% 90.1% 89.1% 88.9% 1.49%39 7 7 0.7 1.88 ∗ 107 3.5140% 224.9% 90.1% 89.1% 88.9% 1.49%40 8 6 0.6 9.81 ∗ 106 1.8290% 117.1% 89.7% 89.3% 88.9% 1.49%41 8 6 0.6 9.99 ∗ 106 1.8620% 119.1% 89.7% 89.3% 88.9% 1.49%42 8 6 0.6 1.01 ∗ 107 1.8940% 121.2% 89.7% 89.3% 88.9% 1.49%43 8 6 0.6 1.03 ∗ 107 1.9270% 123.3% 89.8% 89.4% 88.9% 1.49%44 8 6 0.6 1.05 ∗ 107 1.9590% 125.4% 89.8% 89.5% 89.0% 1.42%45 8 6 0.6 1.06 ∗ 107 1.9920% 127.5% 89.8% 89.5% 89.0% 1.42%46 8 6 0.6 1.08 ∗ 107 2.0240% 129.6% 89.8% 89.5% 89.0% 1.42%47 8 6 0.6 1.10 ∗ 107 2.0570% 131.6% 89.8% 89.5% 89.0% 1.42%48 8 6 0.6 1.12 ∗ 107 2.0890% 133.7% 89.8% 89.5% 89.0% 1.42%64 8 6 0.6 1.12 ∗ 107 2.0890% 133.7% 89.8% 89.5% 89.0% 1.42%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML ω256.

Table A.11: Best achieved PSML Bloom filter TP0 with corresponding TP0.1 and TP0.5, with a finalbinary PSML comparison and tree pre-selection of t = 1, . . . , 48 trees at T = 64.

144

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 4 5 0.4 2.02 ∗ 106 0.3772% 24.14% 52.3% 52.2% 51.9% 38.5%2 5 5 0.8 1.87 ∗ 106 0.3499% 22.39% 62.7% 62.5% 62.5% 27.9%3 5 5 0.8 2.12 ∗ 106 0.3967% 25.39% 70.0% 69.8% 69.8% 20.6%4 5 5 0.8 2.38 ∗ 106 0.4436% 28.39% 74.3% 74.0% 73.9% 16.5%5 4 5 0.7 3.28 ∗ 106 0.6116% 39.14% 77.0% 77.0% 76.4% 14.0%6 4 4 0.6 2.26 ∗ 106 0.4211% 26.95% 79.1% 78.5% 77.5% 12.9%7 6 5 0.6 2.61 ∗ 106 0.4879% 31.22% 80.6% 80.2% 79.6% 10.8%8 4 5 0.6 4.22 ∗ 106 0.7874% 50.39% 83.3% 82.8% 82.0% 8.45%9 4 5 0.6 4.54 ∗ 106 0.8459% 54.14% 84.6% 84.1% 83.2% 7.20%10 4 5 0.6 4.85 ∗ 106 0.9045% 57.89% 85.7% 85.0% 84.1% 6.34%11 4 5 0.6 5.17 ∗ 106 0.9631% 61.64% 86.4% 85.6% 84.7% 5.71%12 4 5 0.6 5.48 ∗ 106 1.0220% 65.39% 86.6% 85.9% 85.0% 5.40%13 4 5 0.6 5.79 ∗ 106 1.0800% 69.14% 87.4% 86.7% 85.8% 4.62%14 4 5 0.6 6.11 ∗ 106 1.1390% 72.89% 88.0% 87.3% 86.4% 3.99%15 4 5 0.6 6.42 ∗ 106 1.1980% 76.64% 88.4% 87.7% 86.7% 3.68%16 4 5 0.6 6.74 ∗ 106 1.2560% 80.39% 88.6% 88.0% 87.0% 3.45%17 4 5 0.6 7.05 ∗ 106 1.3150% 84.14% 88.7% 88.0% 87.0% 3.37%18 5 7 0.6 1.68 ∗ 107 3.1310% 200.4% 87.9% 87.5% 86.8% 3.60%19 4 7 0.6 2.19 ∗ 107 4.0800% 261.1% 88.5% 87.8% 87.0% 3.37%20 4 7 0.6 2.28 ∗ 107 4.2470% 271.8% 88.7% 88.0% 87.2% 3.21%21 4 7 0.6 2.37 ∗ 107 4.4150% 282.5% 88.6% 88.2% 87.4% 2.98%22 8 5 0.8 4.33 ∗ 106 0.8069% 51.64% 88.3% 87.7% 87.5% 2.90%23 5 7 0.6 2.04 ∗ 107 3.8010% 243.2% 88.7% 88.4% 87.6% 2.82%24 7 4 0.7 3.32 ∗ 106 0.6199% 39.68% 88.8% 87.9% 87.8% 2.59%25 7 4 0.7 3.44 ∗ 106 0.6409% 41.02% 88.8% 87.9% 87.8% 2.59%26 7 4 0.7 3.55 ∗ 106 0.6618% 42.35% 88.8% 88.0% 87.9% 2.51%32 7 4 0.7 3.62 ∗ 106 0.6757% 43.25% 89.0% 88.1% 88.0% 2.35%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML ω256.

Table A.12: Best achieved PSML Bloom filter TP0 with corresponding TP0.1 and TP0.5, with a finalbinary PSML comparison and tree pre-selection of t = 1, . . . , 28 trees at T = 32.

145

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 4 5 0.1 3.59 ∗ 106 1.340% 42.89% 72.7% 72.3% 72.3% 18.1%2 4 5 0.2 3.80 ∗ 106 1.418% 45.39% 80.2% 79.9% 79.6% 10.8%3 4 4 0.6 2.52 ∗ 106 0.940% 30.08% 83.4% 82.7% 82.7% 7.67%4 4 5 0.4 4.22 ∗ 106 1.575% 50.39% 85.2% 84.8% 84.5% 5.95%5 4 5 0.5 4.43 ∗ 106 1.653% 52.89% 86.9% 85.9% 85.8% 4.62%6 4 5 0.6 4.64 ∗ 106 1.731% 55.39% 87.2% 86.5% 86.5% 3.92%7 5 5 0.4 3.89 ∗ 106 1.450% 46.39% 88.4% 87.5% 87.4% 2.98%8 4 4 0.6 3.17 ∗ 106 1.184% 37.89% 89.0% 88.2% 88.2% 2.2%9 4 4 0.6 3.30 ∗ 106 1.233% 39.45% 89.4% 88.6% 88.6% 1.81%10 4 4 0.6 3.44 ∗ 106 1.282% 41.02% 89.8% 89.1% 88.8% 1.57%11 5 4 0.6 2.86 ∗ 106 1.067% 34.14% 89.8% 88.9% 88.9% 1.49%12 4 4 0.6 3.70 ∗ 106 1.379% 44.14% 90.2% 89.4% 89.1% 1.26%13 4 5 0.2 6.11 ∗ 106 2.278% 72.89% 90.1% 89.5% 89.3% 1.1%14 4 5 0.2 6.32 ∗ 106 2.356% 75.39% 90.2% 89.7% 89.5% 0.95%15 4 4 0.6 4.09 ∗ 106 1.526% 48.83% 90.6% 89.7% 89.5% 0.95%16 4 4 0.6 4.22 ∗ 106 1.575% 50.39% 90.6% 89.7% 89.5% 0.95%17 5 4 0.6 3.49 ∗ 106 1.301% 41.64% 90.5% 89.6% 89.6% 0.79%18 6 4 0.6 3.00 ∗ 106 1.119% 35.81% 90.6% 89.7% 89.7% 0.71%19 7 4 0.4 2.65 ∗ 106 0.989% 31.64% 90.6% 89.8% 89.8% 0.63%20 7 4 0.4 2.72 ∗ 106 1.017% 32.53% 90.7% 89.9% 89.8% 0.56%21 7 4 0.4 2.80 ∗ 106 1.045% 33.43% 90.8% 90.0% 89.9% 0.48%22 7 4 0.4 2.87 ∗ 106 1.072% 34.32% 90.9% 90.1% 90.0% 0.4%23 7 4 0.4 2.95 ∗ 106 1.100% 35.21% 90.9% 90.2% 90.1% 0.32%24 7 4 0.4 3.02 ∗ 106 1.128% 36.10% 91.1% 90.3% 90.2% 0.17%25 6 6 0.1 9.58 ∗ 106 3.571% 114.3% 91.1% 90.5% 90.4% 0.0%26 6 6 0.1 9.81 ∗ 106 3.658% 117.1% 91.1% 90.5% 90.4% 0.0%27 6 6 0.1 1.00 ∗ 107 3.745% 119.8% 90.6% 90.5% 90.4% 0.0%28 6 6 0.1 1.02 ∗ 107 3.832% 122.6% 90.5% 90.5% 90.3% 0.1%29 6 6 0.1 1.05 ∗ 107 3.918% 125.4% 90.5% 90.5% 90.3% 0.1%30 6 6 0.1 1.07 ∗ 107 4.005% 128.2% 90.5% 90.5% 90.3% 0.1%31 6 6 0.1 1.09 ∗ 107 4.092% 130.9% 90.5% 90.5% 90.3% 0.1%32 6 6 0.1 1.12 ∗ 107 4.179% 133.7% 90.5% 90.5% 90.3% 0.1%33 6 6 0.1 1.14 ∗ 107 4.266% 136.5% 90.5% 90.5% 90.3% 0.1%34 6 6 0.1 1.16 ∗ 107 4.352% 139.3% 90.5% 90.5% 90.3% 0.1%35 4 6 0.1 1.78 ∗ 107 6.653% 212.9% 90.9% 90.7% 90.5% +0.1%36 4 6 0.1 1.82 ∗ 107 6.783% 217.1% 90.9% 90.7% 90.5% +0.1%37 4 6 0.1 1.85 ∗ 107 6.913% 221.2% 90.9% 90.7% 90.5% +0.1%38 4 6 0.1 1.89 ∗ 107 7.043% 225.4% 91.0% 90.8% 90.5% +0.1%39 4 6 0.1 1.92 ∗ 107 7.174% 229.6% 91.0% 90.8% 90.5% +0.1%40 4 6 0.1 1.96 ∗ 107 7.304% 233.7% 91.0% 90.8% 90.5% +0.1%41 4 6 0.1 1.99 ∗ 107 7.434% 237.9% 91.0% 90.8% 90.5% +0.1%42 4 6 0.1 2.03 ∗ 107 7.564% 242.1% 91.0% 90.8% 90.5% +0.1%43 4 6 0.1 2.06 ∗ 107 7.694% 246.2% 91.0% 90.8% 90.5% +0.1%44 4 6 0.1 2.10 ∗ 107 7.825% 250.4% 91.0% 90.8% 90.5% +0.1%45 4 6 0.1 2.13 ∗ 107 7.955% 254.6% 91.0% 90.8% 90.5% +0.1%46 4 6 0.1 2.17 ∗ 107 8.085% 258.7% 91.0% 90.8% 90.5% +0.1%47 4 6 0.1 2.20 ∗ 107 8.215% 262.9% 91.0% 90.8% 90.5% +0.1%48 4 6 0.1 2.24 ∗ 107 8.346% 267.1% 91.0% 90.8% 90.5% +0.1%64 4 6 0.1 2.24 ∗ 107 8.346% 267.1% 91.0% 90.8% 90.5% +0.1%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML ω256.

Table A.13: Best achieved binary PSML Bloom filter TP0 with corresponding TP0.1 and TP0.5, witha final real-valued PSML comparison and tree pre-selection of t = 1, . . . , 48 trees at T = 64.

146

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.47 0.64 0.65 3.48 ∗ 105 0.1297% 26.56% 49.61% 49.3% 48.8% 48.6% 41.8%2 0.49 0.63 0.65 3.69 ∗ 105 0.1373% 28.12% 37.27% 61.0% 60.5% 60.4% 30.0%3 0.49 0.63 0.65 3.89 ∗ 105 0.1450% 29.69% 29.14% 68.9% 68.3% 68.1% 22.3%4 0.46 0.60 0.65 4.10 ∗ 105 0.1526% 31.25% 25.55% 72.4% 72.0% 72.0% 18.4%5 0.46 0.60 0.65 4.30 ∗ 105 0.1602% 32.81% 22.27% 75.5% 75.3% 74.6% 15.8%6 0.46 0.63 0.65 4.51 ∗ 105 0.1678% 34.38% 18.98% 78.0% 77.7% 77.3% 13.1%7 0.46 0.60 0.65 4.71 ∗ 105 0.1755% 35.94% 16.72% 80.6% 80.0% 79.7% 10.7%8 0.46 0.60 0.65 4.92 ∗ 105 0.1831% 37.50% 14.84% 82.1% 81.5% 81.2% 9.20%9 0.46 0.60 0.65 5.12 ∗ 105 0.1907% 39.06% 13.83% 83.1% 82.3% 82.0% 8.37%10 0.46 0.60 0.65 5.32 ∗ 105 0.1984% 40.62% 12.81% 83.9% 83.4% 83.0% 7.35%11 0.46 0.60 0.65 5.53 ∗ 105 0.2060% 42.19% 11.80% 84.5% 84.1% 83.8% 6.57%12 0.46 0.60 0.65 5.73 ∗ 105 0.2136% 43.75% 10.94% 85.3% 84.9% 84.6% 5.79%13 0.46 0.60 0.65 5.94 ∗ 105 0.2213% 45.31% 10.00% 86.0% 85.5% 85.2% 5.24%14 0.46 0.60 0.65 6.14 ∗ 105 0.2289% 46.88% 9.375% 86.4% 85.9% 85.5% 4.85%15 0.46 0.60 0.65 6.35 ∗ 105 0.2365% 48.44% 8.750% 87.0% 86.5% 86.1% 4.31%16 0.46 0.60 0.65 6.55 ∗ 105 0.2441% 50.00% 8.203% 87.4% 86.9% 86.5% 3.92%17 0.46 0.63 0.65 6.76 ∗ 105 0.2518% 51.56% 7.656% 87.8% 87.5% 87.0% 3.45%18 0.46 0.63 0.65 6.96 ∗ 105 0.2594% 53.12% 7.109% 88.3% 88.0% 87.4% 2.98%19 0.46 0.63 0.65 7.17 ∗ 105 0.2670% 54.69% 6.719% 88.6% 88.2% 87.7% 2.74%20 0.46 0.63 0.65 7.37 ∗ 105 0.2747% 56.25% 6.406% 89.0% 88.3% 88.0% 2.35%21 0.46 0.63 0.65 7.58 ∗ 105 0.2823% 57.81% 6.094% 89.1% 88.4% 88.1% 2.28%22 0.49 0.68 0.65 7.78 ∗ 105 0.2899% 59.38% 6.016% 88.8% 88.4% 88.4% 2.04%23 0.49 0.68 0.65 7.99 ∗ 105 0.2975% 60.94% 5.859% 89.1% 88.6% 88.5% 1.88%24 0.46 0.62 0.70 8.19 ∗ 105 0.3052% 62.50% 5.547% 89.5% 88.8% 88.8% 1.65%25 0.46 0.62 0.70 8.40 ∗ 105 0.3128% 64.06% 5.469% 89.5% 88.8% 88.8% 1.57%26 0.46 0.62 0.65 8.60 ∗ 105 0.3204% 65.62% 5.078% 89.5% 89.1% 88.8% 1.57%27 0.46 0.62 0.70 8.81 ∗ 105 0.3281% 67.19% 4.922% 89.8% 89.1% 89.1% 1.26%28 0.46 0.62 0.70 9.01 ∗ 105 0.3357% 68.75% 4.766% 90.0% 89.2% 89.2% 1.18%29 0.46 0.62 0.70 9.22 ∗ 105 0.3433% 70.31% 4.766% 90.0% 89.2% 89.2% 1.18%30 0.46 0.62 0.70 9.42 ∗ 105 0.3510% 71.88% 4.688% 90.1% 89.3% 89.3% 1.10%31 0.46 0.62 0.70 9.63 ∗ 105 0.3586% 73.44% 4.688% 90.1% 89.3% 89.3% 1.10%32 0.46 0.62 0.70 9.83 ∗ 105 0.3662% 75.00% 4.531% 90.2% 89.4% 89.4% 1.03%33 0.46 0.62 0.65 1.00 ∗ 106 0.3738% 76.56% 4.062% 89.8% 89.7% 89.5% 0.95%34 0.46 0.62 0.65 1.02 ∗ 106 0.3815% 78.12% 4.062% 89.8% 89.7% 89.5% 0.95%35 0.46 0.62 0.70 1.04 ∗ 106 0.3891% 79.69% 4.375% 89.9% 89.5% 89.5% 0.87%36 0.46 0.62 0.70 1.06 ∗ 106 0.3967% 81.25% 4.375% 89.9% 89.5% 89.5% 0.87%37 0.46 0.62 0.70 1.09 ∗ 106 0.4044% 82.81% 4.375% 89.8% 89.5% 89.5% 0.87%38 0.46 0.62 0.70 1.11 ∗ 106 0.4102% 84.38% 4.141% 89.9% 89.7% 89.7% 0.71%39 0.46 0.62 0.70 1.13 ∗ 106 0.4196% 85.94% 4.141% 89.9% 89.7% 89.7% 0.71%40 0.46 0.62 0.70 1.15 ∗ 106 0.4272% 87.50% 3.984% 90.0% 89.8% 89.8% 0.63%41 0.46 0.62 0.70 1.17 ∗ 106 0.4349% 89.06% 3.984% 90.0% 89.8% 89.8% 0.63%42 0.46 0.62 0.70 1.19 ∗ 106 0.4425% 90.62% 3.984% 90.0% 89.8% 89.8% 0.63%43 0.46 0.62 0.70 1.21 ∗ 106 0.4501% 92.19% 3.906% 90.0% 89.8% 89.8% 0.63%44 0.46 0.62 0.70 1.23 ∗ 106 0.4578% 93.75% 3.906% 90.0% 89.8% 89.8% 0.63%45 0.46 0.62 0.70 1.25 ∗ 106 0.4654% 95.31% 3.906% 90.0% 89.8% 89.8% 0.63%46 0.46 0.62 0.70 1.27 ∗ 106 0.4730% 96.88% 3.828% 90.0% 89.8% 89.8% 0.63%47 0.46 0.62 0.70 1.29 ∗ 106 0.4807% 98.44% 3.750% 90.1% 89.8% 89.8% 0.56%48 0.46 0.62 0.70 1.31 ∗ 106 0.4883% 100.0% 3.750% 90.1% 89.8% 89.8% 0.56%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table A.14: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using SMR-CPCA-C templates at T = 64.

147

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.64 0.65 3.48 ∗ 105 0.1297% 26.56% 28.75% 70.2% 69.5% 69.3% 21.1%2 0.49 0.64 0.65 3.69 ∗ 105 0.1373% 28.12% 18.44% 79.0% 78.4% 78.0% 12.4%3 0.49 0.67 0.70 3.89 ∗ 105 0.1450% 29.69% 14.22% 83.0% 82.7% 81.9% 8.53%4 0.46 0.62 0.65 4.10 ∗ 105 0.1526% 31.25% 11.88% 84.5% 84.2% 83.9% 6.49%5 0.48 0.66 0.70 4.30 ∗ 105 0.1602% 32.81% 10.00% 86.6% 85.8% 85.5% 4.93%6 0.48 0.66 0.70 4.51 ∗ 105 0.1678% 34.38% 8.516% 87.2% 86.5% 86.3% 4.07%7 0.49 0.64 0.70 4.71 ∗ 105 0.1755% 35.94% 7.734% 88.0% 87.3% 87.0% 3.37%8 0.49 0.68 0.70 4.92 ∗ 105 0.1831% 37.50% 6.875% 88.8% 87.8% 87.4% 2.98%9 0.46 0.62 0.65 5.12 ∗ 105 0.1907% 39.06% 7.266% 88.1% 88.0% 87.8% 2.59%10 0.49 0.68 0.70 5.32 ∗ 105 0.1984% 40.62% 5.859% 89.1% 88.6% 88.1% 2.28%11 0.49 0.68 0.70 5.53 ∗ 105 0.2060% 42.19% 5.391% 89.5% 88.9% 88.4% 2.04%12 0.49 0.68 0.70 5.73 ∗ 105 0.2136% 43.75% 5.312% 89.5% 88.9% 88.4% 2.04%13 0.45 0.78 0.65 5.94 ∗ 105 0.2213% 45.31% 5.703% 89.5% 88.8% 88.7% 1.73%14 0.45 0.78 0.65 6.14 ∗ 105 0.2289% 46.88% 5.469% 89.6% 88.8% 88.8% 1.65%15 0.47 0.61 0.65 6.35 ∗ 105 0.2365% 48.44% 5.312% 89.2% 89.1% 88.9% 1.49%16 0.47 0.61 0.65 6.55 ∗ 105 0.2441% 50.00% 5.078% 89.3% 89.2% 89.0% 1.42%17 0.46 0.62 0.65 6.76 ∗ 105 0.2518% 51.56% 4.453% 89.4% 89.4% 89.1% 1.26%18 0.46 0.62 0.65 6.96 ∗ 105 0.2594% 53.12% 4.375% 89.5% 89.5% 89.3% 1.10%19 0.46 0.62 0.65 7.17 ∗ 105 0.2670% 54.69% 4.297% 89.6% 89.6% 89.4% 1.03%20 0.46 0.62 0.65 7.37 ∗ 105 0.2747% 56.25% 4.375% 89.6% 89.6% 89.4% 1.03%21 0.46 0.62 0.65 7.58 ∗ 105 0.2823% 57.81% 4.219% 89.8% 89.8% 89.5% 0.87%22 0.46 0.62 0.65 7.78 ∗ 105 0.2899% 59.38% 4.141% 89.8% 89.8% 89.6% 0.79%23 0.46 0.62 0.65 7.99 ∗ 105 0.2975% 60.94% 3.906% 89.9% 89.9% 89.7% 0.71%24 0.46 0.62 0.65 8.19 ∗ 105 0.3052% 62.50% 3.906% 89.9% 89.9% 89.7% 0.71%25 0.46 0.62 0.65 8.40 ∗ 105 0.3128% 64.06% 3.906% 89.9% 89.9% 89.7% 0.71%26 0.46 0.62 0.65 8.60 ∗ 105 0.3204% 65.62% 3.828% 89.9% 89.9% 89.7% 0.71%27 0.46 0.62 0.65 8.81 ∗ 105 0.3281% 67.19% 3.828% 89.9% 89.9% 89.7% 0.71%28 0.46 0.62 0.65 9.01 ∗ 105 0.3357% 68.75% 3.750% 90.0% 90.0% 89.8% 0.63%29 0.46 0.62 0.65 9.22 ∗ 105 0.3433% 70.31% 3.750% 90.0% 90.0% 89.8% 0.63%30 0.46 0.62 0.65 9.42 ∗ 105 0.3510% 71.88% 3.750% 90.0% 90.0% 89.8% 0.63%31 0.46 0.62 0.65 9.63 ∗ 105 0.3586% 73.44% 3.672% 90.0% 90.0% 89.8% 0.63%32 0.46 0.62 0.65 9.83 ∗ 105 0.3662% 75.00% 3.672% 90.0% 90.0% 89.8% 0.63%33 0.46 0.62 0.65 1.00 ∗ 106 0.3738% 76.56% 3.672% 90.0% 90.0% 89.8% 0.63%34 0.46 0.62 0.65 1.02 ∗ 106 0.3815% 78.12% 3.672% 90.0% 90.0% 89.8% 0.63%35 0.46 0.62 0.65 1.04 ∗ 106 0.3891% 79.69% 3.672% 90.0% 90.0% 89.8% 0.63%36 0.46 0.62 0.65 1.06 ∗ 106 0.3967% 81.25% 3.594% 90.0% 90.0% 89.8% 0.63%37 0.46 0.62 0.65 1.09 ∗ 106 0.4044% 82.81% 3.594% 90.0% 90.0% 89.8% 0.63%38 0.46 0.62 0.65 1.11 ∗ 106 0.4120% 84.38% 3.594% 90.0% 90.0% 89.8% 0.63%39 0.46 0.62 0.65 1.13 ∗ 106 0.4196% 85.94% 3.594% 90.0% 90.0% 89.8% 0.63%40 0.46 0.62 0.65 1.15 ∗ 106 0.4272% 87.50% 3.594% 90.0% 90.0% 89.8% 0.63%41 0.46 0.62 0.65 1.17 ∗ 106 0.4349% 89.06% 3.594% 90.0% 90.0% 89.8% 0.63%42 0.46 0.62 0.65 1.19 ∗ 106 0.4425% 90.62% 3.594% 90.0% 90.0% 89.8% 0.63%43 0.46 0.62 0.65 1.21 ∗ 106 0.4501% 92.19% 3.672% 90.0% 90.0% 89.8% 0.63%44 0.46 0.62 0.65 1.23 ∗ 106 0.4578% 93.75% 3.672% 90.0% 90.0% 89.8% 0.63%45 0.46 0.62 0.65 1.25 ∗ 106 0.4654% 95.31% 3.672% 90.0% 90.0% 89.8% 0.63%46 0.46 0.62 0.65 1.27 ∗ 106 0.4730% 96.88% 3.672% 90.0% 90.0% 89.8% 0.63%47 0.46 0.62 0.65 1.29 ∗ 106 0.4807% 98.44% 3.672% 90.0% 90.0% 89.8% 0.63%48 0.46 0.62 0.65 1.31 ∗ 106 0.4883% 100.0% 3.672% 90.0% 90.0% 89.8% 0.63%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table A.15: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using SMR-CPCA-M templates at T = 64.

148

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.60 0.70 1.95 ∗ 105 0.0725% 14.84% 55.23% 43.9% 43.8% 43.7% 46.7%2 0.49 0.60 0.75 2.25 ∗ 105 0.0839% 17.19% 41.25% 57.4% 57.3% 57.0% 33.4%3 0.49 0.60 0.75 2.56 ∗ 105 0.0954% 19.53% 33.20% 65.2% 64.5% 64.5% 25.9%4 0.46 0.63 0.70 2.87 ∗ 105 0.1068% 21.88% 26.88% 71.1% 70.0% 69.8% 20.6%5 0.49 0.68 0.75 3.17 ∗ 105 0.1183% 24.22% 22.73% 74.6% 73.8% 73.6% 16.8%6 0.49 0.68 0.75 3.48 ∗ 105 0.1297% 26.56% 20.00% 77.0% 75.9% 75.9% 14.5%7 0.49 0.68 0.70 3.79 ∗ 105 0.1411% 28.91% 17.34% 79.2% 78.3% 78.1% 12.3%8 0.47 0.64 0.75 4.10 ∗ 105 0.1526% 31.25% 15.23% 81.2% 80.9% 79.9% 10.5%9 0.49 0.68 0.70 4.40 ∗ 105 0.1640% 33.59% 13.44% 82.7% 81.8% 81.2% 9.20%10 0.46 0.79 0.65 4.71 ∗ 105 0.1755% 35.94% 14.22% 82.6% 82.3% 82.3% 8.13%11 0.48 0.66 0.70 5.02 ∗ 105 0.1869% 38.28% 11.33% 84.5% 83.7% 83.3% 7.12%12 0.48 0.66 0.70 5.32 ∗ 105 0.1984% 40.62% 10.39% 85.3% 84.5% 84.1% 5.34%13 0.48 0.66 0.70 5.63 ∗ 105 0.2098% 42.97% 9.688% 85.9% 84.9% 84.7% 5.71%14 0.48 0.66 0.70 5.94 ∗ 105 0.2213% 45.31% 9.219% 86.2% 85.9% 85.3% 5.09%15 0.49 0.67 0.70 6.25 ∗ 105 0.2327% 47.66% 8.281% 87.3% 87.0% 85.9% 4.46%16 0.46 0.62 0.65 6.55 ∗ 105 0.2441% 50.00% 8.359% 86.7% 86.6% 86.5% 3.92%17 0.46 0.62 0.65 6.86 ∗ 105 0.2556% 52.34% 7.891% 87.1% 87.1% 87.0% 3.45%18 0.46 0.62 0.65 7.17 ∗ 105 0.2670% 54.69% 7.422% 87.4% 87.4% 87.3% 3.13%19 0.46 0.62 0.65 7.48 ∗ 105 0.2785% 57.03% 7.344% 87.5% 87.5% 87.3% 3.06%20 0.46 0.62 0.65 7.78 ∗ 105 0.2899% 59.38% 7.109% 87.7% 87.7% 87.6% 2.82%21 0.46 0.62 0.65 8.09 ∗ 105 0.3014% 61.72% 6.719% 88.1% 88.1% 88.0% 2.43%22 0.46 0.62 0.65 8.40 ∗ 105 0.3128% 64.06% 6.484% 88.3% 88.3% 88.1% 2.28%23 0.46 0.62 0.65 8.70 ∗ 105 0.3242% 66.41% 6.406% 88.4% 88.4% 88.2% 2.20%24 0.46 0.62 0.65 9.01 ∗ 105 0.3357% 68.75% 6.016% 88.7% 88.7% 88.5% 1.88%25 0.46 0.62 0.65 9.32 ∗ 105 0.3471% 71.09% 5.938% 88.7% 88.7% 88.5% 1.88%26 0.46 0.62 0.65 9.63 ∗ 105 0.3586% 73.44% 5.625% 88.8% 88.8% 88.7% 1.73%27 0.46 0.62 0.65 9.93 ∗ 105 0.3700% 75.78% 5.469% 88.8% 88.8% 88.7% 1.73%28 0.46 0.62 0.65 1.02 ∗ 106 0.3815% 78.12% 5.469% 88.8% 88.8% 88.7% 1.73%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table A.16: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 26 trees using SMR-CPCA-M templates at T = 32.

149

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.64 0.65 3.53 ∗ 105 0.1316% 26.95% 28.75% 70.2% 69.5% 69.3% 21.1%2 0.49 0.64 0.65 3.74 ∗ 105 0.1392% 28.52% 18.44% 79% 78.4% 78% 12.4%3 0.49 0.67 0.70 3.94 ∗ 105 0.1469% 30.08% 14.22% 83% 82.7% 81.9% 8.53%4 0.46 0.62 0.65 4.15 ∗ 105 0.1545% 31.64% 11.88% 84.5% 84.2% 83.9% 6.49%5 0.48 0.66 0.70 4.35 ∗ 105 0.1621% 33.20% 10.0% 86.6% 85.8% 85.5% 4.93%6 0.48 0.66 0.70 4.56 ∗ 105 0.1698% 34.77% 8.516% 87.2% 86.5% 86.3% 4.07%7 0.49 0.64 0.70 4.76 ∗ 105 0.1774% 36.33% 7.734% 88% 87.3% 87% 3.37%8 0.49 0.68 0.70 4.97 ∗ 105 0.1850% 37.89% 6.875% 88.8% 87.8% 87.4% 2.98%9 0.46 0.62 0.70 5.17 ∗ 105 0.1926% 39.45% 7.266% 88.1% 88% 87.8% 2.59%10 0.49 0.68 0.70 5.38 ∗ 105 0.2003% 41.02% 5.859% 89.1% 88.6% 88.1% 2.28%11 0.49 0.68 0.70 5.58 ∗ 105 0.2079% 42.58% 5.391% 89.5% 88.9% 88.4% 2.04%12 0.49 0.68 0.70 5.79 ∗ 105 0.2155% 44.14% 5.312% 89.5% 88.9% 88.4% 2.04%13 0.48 0.65 0.65 5.99 ∗ 105 0.2232% 45.70% 5.156% 89.5% 88.8% 88.4% 1.96%14 0.48 0.65 0.65 6.20 ∗ 105 0.2308% 47.27% 4.844% 89.8% 89.1% 88.7% 1.73%15 0.47 0.61 0.65 6.40 ∗ 105 0.2384% 48.83% 5.312% 89.2% 89.1% 88.9% 1.49%16 0.47 0.61 0.65 6.60 ∗ 105 0.2460% 50.39% 5.078% 89.3% 89.2% 89% 1.42%17 0.46 0.62 0.65 6.81 ∗ 105 0.2537% 51.95% 4.453% 89.4% 89.4% 89.1% 1.26%18 0.46 0.62 0.65 7.01 ∗ 105 0.2613% 53.52% 4.375% 89.5% 89.5% 89.3% 1.1%19 0.46 0.62 0.65 7.22 ∗ 105 0.2689% 55.08% 4.297% 89.6% 89.6% 89.4% 1.03%20 0.46 0.62 0.65 7.42 ∗ 105 0.2766% 56.64% 4.375% 89.6% 89.6% 89.4% 1.03%21 0.46 0.62 0.65 7.63 ∗ 105 0.2842% 58.20% 4.219% 89.8% 89.8% 89.5% 0.87%22 0.46 0.62 0.65 7.83 ∗ 105 0.2918% 59.77% 4.141% 89.8% 89.8% 89.6% 0.79%23 0.46 0.62 0.65 8.04 ∗ 105 0.2995% 61.33% 3.906% 89.9% 89.9% 89.7% 0.71%24 0.46 0.62 0.65 8.24 ∗ 105 0.3071% 62.89% 3.906% 89.9% 89.9% 89.7% 0.71%25 0.46 0.62 0.65 8.45 ∗ 105 0.3147% 64.45% 3.906% 89.9% 89.9% 89.7% 0.71%26 0.46 0.62 0.65 8.65 ∗ 105 0.3223% 65.02% 3.828% 89.9% 89.9% 89.7% 0.71%27 0.46 0.62 0.65 8.86 ∗ 105 0.3300% 67.58% 3.828% 89.9% 89.9% 89.7% 0.71%28 0.46 0.62 0.65 9.06 ∗ 105 0.3376% 69.14% 3.75% 90% 90% 89.8% 0.63%29 0.46 0.62 0.65 9.27 ∗ 105 0.3452% 70.70% 3.75% 90% 90% 89.8% 0.63%30 0.46 0.62 0.65 9.47 ∗ 105 0.3529% 72.27% 3.75% 90% 90% 89.8% 0.63%31 0.46 0.62 0.65 9.68 ∗ 105 0.3605% 73.83% 3.672% 90% 90% 89.8% 0.63%32 0.46 0.62 0.65 9.88 ∗ 105 0.3681% 75.39% 3.672% 90% 90% 89.8% 0.63%33 0.46 0.62 0.65 1.01 ∗ 106 0.3757% 76.95% 3.672% 90% 90% 89.8% 0.63%34 0.46 0.62 0.65 1.03 ∗ 106 0.3834% 78.52% 3.672% 90% 90% 89.8% 0.63%35 0.46 0.62 0.65 1.05 ∗ 106 0.3910% 80.08% 3.672% 90% 90% 89.8% 0.63%36 0.46 0.62 0.65 1.07 ∗ 106 0.3986% 81.64% 3.594% 90% 90% 89.8% 0.63%37 0.46 0.62 0.65 1.09 ∗ 106 0.4063% 83.20% 3.594% 90% 90% 89.8% 0.63%38 0.46 0.62 0.65 1.11 ∗ 106 0.4139% 84.77% 3.594% 90% 90% 89.8% 0.63%39 0.46 0.62 0.65 1.13 ∗ 106 0.4215% 86.33% 3.594% 90% 90% 89.8% 0.63%40 0.46 0.62 0.65 1.15 ∗ 106 0.4292% 87.89% 3.594% 90% 90% 89.8% 0.63%41 0.46 0.62 0.65 1.17 ∗ 106 0.4368% 89.45% 3.594% 90% 90% 89.8% 0.63%42 0.46 0.62 0.65 1.19 ∗ 106 0.4444% 91.02% 3.594% 90% 90% 89.8% 0.63%43 0.46 0.62 0.65 1.21 ∗ 106 0.4520% 92.58% 3.672% 90% 90% 89.8% 0.63%44 0.46 0.62 0.65 1.23 ∗ 106 0.4597% 94.14% 3.672% 90% 90% 89.8% 0.63%45 0.46 0.62 0.65 1.25 ∗ 106 0.4673% 95.70% 3.672% 90% 90% 89.8% 0.63%46 0.46 0.62 0.65 1.27 ∗ 106 0.4749% 97.27% 3.672% 90% 90% 89.8% 0.63%47 0.46 0.62 0.65 1.30 ∗ 106 0.4826% 98.83% 3.672% 90% 90% 89.8% 0.63%48 0.46 0.62 0.65 1.32 ∗ 106 0.4902% 100.4% 3.672% 90% 90% 89.8% 0.63%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary SMR ω256.

Table A.17: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using SMR-CPCA-M templates at T = 64 with an additional binaryPSML-CPCA comparison.

150

SMR Workload Recognition Performance

t λmax λCP CAmax MT Bit Threshold ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.65 0.65 0.7 3.53 ∗ 105 0.1316% 26.95% 28.91% 70.0% 69.7% 69.3% 21.1%2 0.49 0.64 0.70 0.5 3.74 ∗ 105 0.1392% 28.52% 18.67% 79.1% 78.4% 78.4% 12.0%3 0.49 0.64 0.70 0.6 3.94 ∗ 105 0.1469% 30.08% 14.22% 82.9% 82.1% 82.0% 8.45%4 0.46 0.62 0.65 0.6 4.15 ∗ 105 0.1545% 31.64% 11.95% 84.8% 84.2% 83.9% 6.49%5 0.48 0.66 0.70 0.5 4.35 ∗ 105 0.1621% 33.20% 10.16% 86.4% 86.0% 85.5% 4.85%6 0.48 0.66 0.70 0.5 4.56 ∗ 105 0.1698% 34.77% 8.750% 87.4% 86.9% 86.5% 3.92%7 0.49 0.64 0.70 0.7 4.76 ∗ 105 0.1774% 36.33% 7.734% 88.1% 87.3% 87.1% 3.29%8 0.49 0.68 0.70 0.6 4.97 ∗ 105 0.1850% 37.89% 6.875% 88.9% 87.9% 87.5% 2.90%9 0.49 0.68 0.70 0.6 5.17 ∗ 105 0.1926% 39.45% 6.484% 88.8% 88.3% 87.8% 2.59%10 0.49 0.68 0.70 0.7 5.38 ∗ 105 0.2003% 41.02% 5.938% 89.2% 88.7% 88.2% 2.20%11 0.49 0.68 0.70 0.7 5.58 ∗ 105 0.2079% 42.58% 5.391% 89.5% 89.0% 88.4% 1.96%12 0.49 0.68 0.70 0.5 5.79 ∗ 105 0.2155% 44.14% 5.469% 89.5% 89.1% 88.4% 1.96%13 0.47 0.64 0.65 0.7 5.99 ∗ 105 0.2232% 45.70% 5.078% 89.5% 89.2% 88.5% 1.88%14 0.46 0.63 0.65 0.5 6.20 ∗ 105 0.2308% 47.27% 5.703% 89.8% 89.1% 88.8% 1.57%15 0.46 0.63 0.65 0.5 6.40 ∗ 105 0.2384% 48.83% 5.312% 90.1% 89.3% 89.0% 1.42%16 0.46 0.62 0.70 0.6 6.60 ∗ 105 0.2460% 50.39% 4.844% 89.5% 89.1% 89.1% 1.26%17 0.46 0.63 0.65 0.5 6.81 ∗ 105 0.2537% 51.95% 5.000% 90.3% 89.5% 89.2% 1.18%18 0.46 0.62 0.65 0.9 7.01 ∗ 105 0.2613% 53.52% 4.375% 89.5% 89.5% 89.3% 1.10%19 0.46 0.62 0.70 0.6 7.22 ∗ 105 0.2689% 55.08% 4.453% 89.8% 89.5% 89.5% 0.95%20 0.46 0.62 0.70 0.6 7.42 ∗ 105 0.2766% 56.64% 4.453% 89.8% 89.5% 89.5% 0.95%21 0.46 0.62 0.65 0.8 7.63 ∗ 105 0.2842% 58.20% 4.219% 89.8% 89.8% 89.5% 0.87%22 0.46 0.62 0.65 0.7 7.83 ∗ 105 0.2918% 59.77% 4.219% 89.8% 89.8% 89.6% 0.79%23 0.46 0.62 0.65 0.7 8.04 ∗ 105 0.2995% 61.33% 3.984% 89.9% 89.9% 89.7% 0.71%24 0.46 0.62 0.65 0.8 8.24 ∗ 105 0.3071% 62.89% 3.906% 89.9% 89.9% 89.7% 0.71%25 0.46 0.62 0.70 0.6 8.45 ∗ 105 0.3147% 64.45% 3.984% 90.0% 89.8% 89.8% 0.63%26 0.46 0.62 0.70 0.6 8.65 ∗ 105 0.3223% 66.02% 3.984% 90.0% 89.8% 89.8% 0.63%27 0.46 0.62 0.70 0.5 8.86 ∗ 105 0.3300% 67.58% 3.984% 90.1% 89.8% 89.8% 0.56%28 0.46 0.62 0.70 0.6 9.06 ∗ 105 0.3376% 69.14% 3.828% 90.2% 89.9% 89.9% 0.48%29 0.46 0.62 0.70 0.6 9.27 ∗ 105 0.3452% 70.70% 3.828% 90.2% 89.9% 89.9% 0.48%30 0.46 0.62 0.70 0.5 9.47 ∗ 105 0.3529% 72.27% 3.906% 90.2% 89.9% 89.9% 0.48%31 0.46 0.62 0.70 0.5 9.68 ∗ 105 0.3605% 73.83% 3.906% 90.2% 89.9% 89.9% 0.48%32 0.46 0.62 0.70 0.5 9.88 ∗ 105 0.3681% 75.39% 3.906% 90.2% 89.9% 89.9% 0.48%33 0.46 0.62 0.70 0.5 1.01 ∗ 106 0.3757% 76.95% 3.828% 90.2% 89.9% 89.9% 0.48%34 0.46 0.62 0.70 0.5 1.03 ∗ 106 0.3834% 78.52% 3.828% 90.2% 89.9% 89.9% 0.48%35 0.46 0.62 0.70 0.6 1.05 ∗ 106 0.3910% 80.08% 3.750% 90.2% 89.9% 89.9% 0.48%36 0.46 0.62 0.70 0.5 1.07 ∗ 106 0.3986% 81.64% 3.750% 90.2% 89.9% 89.9% 0.48%37 0.46 0.62 0.70 0.6 1.09 ∗ 106 0.4063% 83.20% 3.672% 90.2% 89.9% 89.9% 0.48%38 0.46 0.62 0.70 0.6 1.11 ∗ 106 0.4139% 84.77% 3.672% 90.2% 89.9% 89.9% 0.48%39 0.46 0.62 0.70 0.5 1.13 ∗ 106 0.4215% 86.33% 3.750% 90.2% 89.9% 89.9% 0.48%40 0.46 0.62 0.70 0.5 1.15 ∗ 106 0.4292% 87.89% 3.750% 90.2% 89.9% 89.9% 0.48%41 0.46 0.62 0.70 0.5 1.17 ∗ 106 0.4368% 89.45% 3.750% 90.2% 89.9% 89.9% 0.48%42 0.46 0.62 0.70 0.5 1.19 ∗ 106 0.4444% 91.02% 3.750% 90.2% 89.9% 89.9% 0.48%43 0.46 0.62 0.70 0.6 1.21 ∗ 106 0.4520% 92.58% 3.672% 90.2% 89.9% 89.9% 0.48%44 0.46 0.62 0.70 0.6 1.23 ∗ 106 0.4597% 94.14% 3.672% 90.2% 89.9% 89.9% 0.48%45 0.46 0.62 0.70 0.6 1.25 ∗ 106 0.4673% 95.70% 3.672% 90.2% 89.9% 89.9% 0.48%46 0.46 0.62 0.70 0.5 1.27 ∗ 106 0.4749% 97.27% 3.750% 90.2% 89.9% 89.9% 0.48%47 0.46 0.62 0.70 0.6 1.30 ∗ 106 0.4826% 98.83% 3.672% 90.2% 89.9% 89.9% 0.48%48 0.46 0.62 0.70 0.6 1.32 ∗ 106 0.4902% 100.4% 3.672% 90.2% 89.9% 89.9% 0.48%64‡ 0.46 0.62 0.70 0.6 1.32 ∗ 106 0.4902% 100.4% 3.672% 90.2% 89.9% 89.9% 0.48%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary SMR ω256.‡ Without tree pre-selection.

Table A.18: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using SMR-CPCA-M templates at T = 64 with an additional binaryPSML-CPCA comparison and masking out most common set bits.

151

SMR Workload Recognition Performance

t λmax λCP CAmax MT Bit Threshold ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.60 0.70 0.9 2.00 ∗ 105 0.0744% 15.23% 55.23% 43.9% 43.8% 43.7% 46.7%2 0.49 0.60 0.75 0.8 2.30 ∗ 105 0.0858% 17.58% 41.25% 57.5% 57.4% 57.0% 33.4%3 0.49 0.60 0.75 0.5 2.61 ∗ 105 0.0973% 19.92% 32.97% 65.4% 64.7% 64.7% 25.7%4 0.46 0.63 0.70 0.6 2.92 ∗ 105 0.1087% 22.27% 26.80% 71.2% 70.1% 69.9% 20.5%5 0.48 0.66 0.75 0.5 3.23 ∗ 105 0.1202% 24.61% 23.28% 74.5% 73.8% 73.6% 16.8%6 0.48 0.66 0.75 0.5 3.53 ∗ 105 0.1316% 26.95% 19.45% 77.9% 77.3% 76.7% 13.7%7 0.48 0.66 0.75 0.5 3.84 ∗ 105 0.1431% 29.30% 17.34% 79.8% 79.1% 78.6% 11.8%8 0.48 0.66 0.75 0.5 4.15 ∗ 105 0.1545% 31.64% 15.55% 81.2% 80.3% 80.0% 10.4%9 0.49 0.68 0.70 0.9 4.45 ∗ 105 0.1659% 33.98% 13.44% 82.7% 81.8% 81.2% 9.20%10 0.48 0.66 0.75 0.5 4.76 ∗ 105 0.1774% 36.33% 12.81% 83.1% 83.0% 82.7% 7.67%11 0.48 0.66 0.70 0.6 5.07 ∗ 105 0.1888% 38.67% 11.56% 84.4% 84.1% 83.5% 6.88%12 0.48 0.66 0.75 0.5 5.38 ∗ 105 0.2003% 41.02% 10.70% 85.1% 84.9% 84.6% 5.79%13 0.48 0.66 0.75 0.5 5.68 ∗ 105 0.2117% 43.36% 10.00% 85.8% 85.6% 85.3% 5.09%14 0.48 0.66 0.75 0.5 5.99 ∗ 105 0.2232% 45.70% 9.531% 86.1% 86.0% 85.7% 4.70%15 0.48 0.66 0.75 0.5 6.30 ∗ 105 0.2346% 48.05% 8.984% 86.5% 86.4% 86.1% 4.31%16 0.48 0.66 0.75 0.5 6.60 ∗ 105 0.2460% 50.39% 8.438% 87.0% 87.0% 86.6% 4.76%17 0.48 0.66 0.75 0.5 6.91 ∗ 105 0.2575% 52.73% 7.969% 87.4% 87.3% 87.0% 3.37%18 0.48 0.66 0.75 0.5 7.22 ∗ 105 0.2689% 55.08% 7.500% 88.0% 87.9% 87.5% 3.90%19 0.48 0.66 0.75 0.5 7.53 ∗ 105 0.2804% 57.42% 7.188% 88.2% 88.1% 87.7% 2.74%20 0.48 0.66 0.75 0.5 7.83 ∗ 105 0.2918% 59.77% 6.953% 88.3% 88.2% 87.7% 2.67%21 0.48 0.66 0.75 0.5 8.14 ∗ 105 0.3033% 62.11% 6.641% 88.6% 88.5% 88.0% 2.35%22 0.48 0.66 0.75 0.5 8.45 ∗ 105 0.3147% 64.45% 6.016% 89.1% 88.7% 88.2% 2.20%23 0.48 0.66 0.75 0.5 8.76 ∗ 105 0.3262% 66.80% 5.859% 89.1% 88.8% 88.3% 2.12%24 0.46 0.62 0.65 0.9 9.06 ∗ 105 0.3376% 69.14% 6.016% 88.7% 88.7% 88.5% 1.88%25 0.48 0.66 0.75 0.5 9.37 ∗ 105 0.3490% 71.48% 5.859% 89.1% 89.1% 88.6% 1.81%26 0.46 0.62 0.65 0.7 9.68 ∗ 105 0.3605% 73.83% 5.703% 88.8% 88.8% 88.7% 1.73%32‡ 0.46 0.62 0.65 0.7 9.88 ∗ 1005 0.3681% 75.39% 5.312% 88.9% 88.9% 88.8% 1.65%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary SMR ω256.‡ Without tree pre-selection.

Table A.19: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 26 trees using SMR-CPCA-M templates at T = 32 with an additional binaryPSML-CPCA comparison and masking out most common set bits.

152

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.64 0.65 6.76 ∗ 105 0.2518% 51.56% 28.67% 70.5% 70.2% 69.8% 20.6%2 0.49 0.64 0.70 6.96 ∗ 105 0.2594% 53.12% 17.81% 80.5% 80.1% 79.9% 10.5%3 0.49 0.64 0.70 7.17 ∗ 105 0.2670% 54.69% 13.98% 84.1% 83.8% 83.6% 6.8%4 0.48 0.65 0.70 7.37 ∗ 105 0.2747% 56.25% 11.80% 85.4% 85.2% 84.5% 5.9%5 0.48 0.66 0.70 7.58 ∗ 105 0.2823% 57.81% 9.844% 87.3% 86.5% 84.5% 5.9%6 0.48 0.66 0.70 7.78 ∗ 105 0.2899% 59.38% 8.516% 87.9% 87.5% 85.5% 4.9%7 0.49 0.68 0.65 7.99 ∗ 105 0.2975% 60.94% 7.578% 89.3% 88.4% 86.4% 4.0%8 0.49 0.68 0.70 8.19 ∗ 105 0.3052% 62.50% 6.875% 89.9% 88.9% 87.0% 3.5%9 0.47 0.65 0.70 8.40 ∗ 105 0.3128% 64.06% 6.094% 89.6% 89.4% 88.8% 1.7%10 0.49 0.68 0.70 8.60 ∗ 105 0.3204% 65.62% 5.781% 90.2% 89.6% 87.5% 2.9%11 0.47 0.60 0.75 8.81 ∗ 105 0.3281% 67.19% 5.312% 90.2% 89.7% 87.5% 2.9%12 0.47 0.60 0.75 9.01 ∗ 105 0.3357% 68.75% 5.156% 90.5% 89.9% 87.6% 2.8%13 0.48 0.66 0.70 9.22 ∗ 105 0.3433% 70.31% 4.844% 90.5% 90.0% 87.7% 2.7%14 0.48 0.66 0.70 9.42 ∗ 105 0.3510% 71.88% 4.609% 90.6% 90.2% 88.0% 2.4%15 0.48 0.65 0.70 9.63 ∗ 105 0.3586% 73.44% 4.219% 90.9% 90.4% 88.1% 2.3%16 0.48 0.65 0.75 9.83 ∗ 105 0.3662% 75.00% 4.062% 90.8% 90.3% 88.1% 2.3%17 0.49 0.67 0.70 1.00 ∗ 106 0.3738% 76.56% 3.906% 90.9% 90.5% 88.1% 2.3%18 0.49 0.67 0.75 1.02 ∗ 106 0.3815% 78.12% 3.828% 91.0% 90.5% 88.2% 2.2%19 0.49 0.67 0.75 1.04 ∗ 106 0.3891% 79.69% 3.828% 91.0% 90.5% 88.2% 2.2%20 0.47 0.66 0.75 1.06 ∗ 106 0.3967% 81.25% 3.672% 91.0% 90.7% 88.4% 2.0%21 0.47 0.64 0.70 1.09 ∗ 106 0.4044% 82.81% 3.594% 91.2% 90.7% 88.4% 2.0%22 0.46 0.62 0.75 1.11 ∗ 106 0.4120% 84.38% 3.516% 91.2% 90.8% 88.5% 1.9%23 0.46 0.62 0.75 1.13 ∗ 106 0.4196% 85.94% 3.516% 91.2% 90.8% 88.5% 1.9%24 0.49 0.68 0.70 1.15 ∗ 106 0.4272% 87.50% 3.516% 91.1% 90.8% 88.4% 2.0%25 0.49 0.68 0.70 1.17 ∗ 106 0.4349% 89.06% 3.516% 91.1% 90.8% 88.4% 2.0%26 0.47 0.60 0.75 1.19 ∗ 106 0.4425% 90.62% 3.438% 91.2% 90.9% 88.5% 1.9%27 0.47 0.63 0.60 1.21 ∗ 106 0.4501% 92.19% 3.438% 91.3% 90.9% 88.5% 1.9%28 0.48 0.64 0.65 1.23 ∗ 106 0.4578% 93.75% 3.359% 91.2% 90.9% 88.5% 1.9%29 0.48 0.64 0.65 1.25 ∗ 106 0.4654% 95.31% 3.359% 91.2% 90.9% 88.5% 1.9%30 0.48 0.64 0.65 1.27 ∗ 106 0.4730% 96.88% 3.359% 91.2% 90.9% 88.5% 1.9%31 0.47 0.64 0.70 1.29 ∗ 106 0.4807% 98.44% 3.359% 91.2% 90.8% 88.5% 1.9%32 0.47 0.64 0.70 1.31 ∗ 106 0.4883% 100.0% 3.359% 91.2% 90.8% 88.5% 1.9%33 0.47 0.60 0.75 1.33 ∗ 106 0.4959% 101.6% 3.281% 91.2% 90.8% 88.5% 1.9%34 0.47 0.60 0.75 1.35 ∗ 106 0.5035% 103.1% 3.281% 91.2% 90.8% 88.5% 1.9%35 0.47 0.60 0.75 1.37 ∗ 106 0.5112% 104.7% 3.281% 91.2% 90.8% 88.5% 1.9%36 0.47 0.60 0.75 1.39 ∗ 106 0.5188% 106.2% 3.281% 91.2% 90.8% 88.5% 1.9%37 0.46 0.70 0.70 1.41 ∗ 106 0.5264% 107.8% 3.281% 91.0% 90.7% 88.4% 2.0%38 0.48 0.62 0.70 1.43 ∗ 106 0.5341% 109.4% 3.281% 91.2% 90.8% 88.5% 1.9%39 0.46 0.70 0.70 1.45 ∗ 106 0.5417% 110.9% 3.203% 91.0% 90.7% 88.4% 2.0%40 0.46 0.70 0.70 1.47 ∗ 106 0.5493% 112.5% 3.203% 91.0% 90.7% 88.4% 2.0%41 0.46 0.70 0.70 1.50 ∗ 106 0.5569% 114.1% 3.203% 91.0% 90.7% 88.4% 2.0%42 0.46 0.70 0.70 1.52 ∗ 106 0.5646% 115.6% 3.203% 91.0% 90.7% 88.4% 2.0%43 0.46 0.70 0.70 1.54 ∗ 106 0.5722% 117.2% 3.203% 91.0% 90.7% 88.4% 2.0%44 0.46 0.70 0.70 1.56 ∗ 106 0.5798% 118.8% 3.125% 91.1% 90.8% 88.4% 2.0%45 0.46 0.70 0.70 1.58 ∗ 106 0.5875% 120.3% 3.125% 91.1% 90.8% 88.4% 2.0%46 0.46 0.70 0.70 1.60 ∗ 106 0.5951% 121.9% 3.125% 91.1% 90.7% 88.4% 2.0%47 0.46 0.70 0.70 1.62 ∗ 106 0.6027% 123.4% 3.125% 91.1% 90.7% 88.4% 2.0%48 0.46 0.70 0.70 1.64 ∗ 106 0.6104% 125.0% 3.125% 91.1% 90.7% 88.4% 2.0%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary SMR ω256.

Table A.20: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using SMR-CPCA-M templates at T = 64 with an additional realPSML-CPCA comparison.

153

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.60 0.70 5.22 ∗ 105 0.1945% 39.84% 55.23% 44.1% 43.8% 43.8% 46.7%2 0.49 0.60 0.75 5.53 ∗ 105 0.206% 42.19% 41.25% 58% 57.7% 57.3% 33.1%3 0.48 0.65 0.75 5.84 ∗ 105 0.2174% 44.53% 32.66% 66% 65.9% 65.6% 24.8%4 0.46 0.63 0.70 6.14 ∗ 105 0.2289% 46.88% 26.88% 71.8% 71.7% 71.2% 19.2%5 0.48 0.67 0.70 6.45 ∗ 105 0.2403% 49.22% 22.66% 75.7% 75.6% 74.5% 15.9%6 0.48 0.66 0.75 6.76 ∗ 105 0.2518% 51.56% 19.45% 78.9% 78.3% 78.1% 12.3%7 0.47 0.64 0.75 7.07 ∗ 105 0.2632% 53.91% 17.27% 80.4% 79.7% 79.1% 11.3%8 0.49 0.68 0.70 7.37 ∗ 105 0.2747% 56.25% 15% 82.5% 82.2% 81.2% 9.2%9 0.49 0.68 0.70 7.68 ∗ 105 0.2861% 58.59% 13.44% 83.8% 83.6% 82.7% 7.74%10 0.47 0.64 0.70 7.99 ∗ 105 0.2975% 60.94% 12.34% 84.4% 83.6% 83.4% 7.04%11 0.47 0.64 0.75 8.29 ∗ 105 0.309% 63.28% 11.25% 85.7% 84.7% 84.5% 5.87%12 0.47 0.64 0.70 8.60 ∗ 105 0.3204% 65.62% 10.16% 86.4% 85.5% 85.4% 5.01%13 0.48 0.65 0.70 8.91 ∗ 105 0.3319% 67.97% 9.141% 87.2% 86.3% 86.2% 4.23%14 0.48 0.65 0.70 9.22 ∗ 105 0.3433% 70.31% 8.438% 87.9% 87% 86.9% 3.53%15 0.48 0.65 0.70 9.52 ∗ 105 0.3548% 72.66% 7.734% 88.5% 87.7% 87.5% 2.9%16 0.47 0.64 0.70 9.83 ∗ 105 0.3662% 75% 7.266% 89.1% 88.4% 88.2% 2.2%17 0.47 0.64 0.70 1.01 ∗ 106 0.3777% 77.34% 6.875% 88.8% 88% 86.9% 3.53%18 0.47 0.64 0.70 1.04 ∗ 106 0.3891% 79.69% 6.641% 89.2% 88.4% 87.2% 3.21%19 0.47 0.64 0.70 1.08 ∗ 106 0.4005% 82.03% 6.25% 89.5% 88.7% 87.5% 2.9%20 0.47 0.64 0.70 1.11 ∗ 106 0.412% 84.38% 5.859% 89.5% 89.1% 86.9% 3.53%21 0.47 0.64 0.70 1.14 ∗ 106 0.4234% 86.72% 5.781% 89.6% 89.1% 87% 3.45%22 0.47 0.64 0.70 1.17 ∗ 106 0.4349% 89.06% 5.625% 89.7% 89.2% 87% 3.45%23 0.48 0.62 0.70 1.20 ∗ 106 0.4463% 91.41% 5.312% 89.5% 89.2% 88.5% 1.88%24 0.49 0.65 0.65 1.23 ∗ 106 0.4578% 93.75% 5.078% 89.8% 89.5% 87.3% 3.06%25 0.49 0.65 0.65 1.26 ∗ 106 0.4692% 96.09% 4.922% 90% 89.6% 87.5% 2.9%26 0.48 0.62 0.65 1.29 ∗ 106 0.4807% 98.44% 4.766% 90.2% 89.8% 87.5% 2.9%27 0.48 0.62 0.70 1.32 ∗ 106 0.4921% 100.8% 4.688% 90.2% 89.8% 87.5% 2.9%28 0.49 0.65 0.65 1.35 ∗ 106 0.5035% 103.1% 4.531% 90.2% 89.8% 87.7% 2.74%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary SMR ω256.

Table A.21: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 26 trees using SMR-CPCA-M templates at T = 32 with an additional realPSML-CPCA comparison.

154

SMR Workload Recognition Performance

t λmax λCP CAmax MT Bit Threshold ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.49 0.64 0.65 0.8 6.76 ∗ 105 0.2518% 51.56% 28.05% 71.3% 70.9% 67.6% 22.8%2 0.49 0.64 0.70 0.9 6.96 ∗ 105 0.2594% 53.12% 17.66% 80.9% 80.4% 73.9% 16.5%3 0.49 0.64 0.70 0.7 7.17 ∗ 105 0.2670% 54.69% 12.81% 84.5% 84.1% 83.4% 7.0%4 0.49 0.64 0.70 0.6 7.37 ∗ 105 0.2747% 56.25% 10.94% 85.7% 85.1% 84.8% 5.6%5 0.48 0.66 0.70 0.9 7.58 ∗ 105 0.2823% 57.81% 9.844% 86.0% 85.8% 85.5% 4.9%6 0.48 0.66 0.70 0.7 7.78 ∗ 105 0.2899% 59.38% 9.141% 86.5% 86.2% 85.8% 4.6%7 0.49 0.68 0.65 0.8 7.99 ∗ 105 0.2975% 60.94% 8.359% 87.3% 87.1% 86.7% 3.7%8 0.49 0.68 0.70 0.6 8.19 ∗ 105 0.3052% 62.50% 8.047% 87.6% 87.3% 86.9% 3.5%9 0.47 0.65 0.70 0.9 8.40 ∗ 105 0.3128% 64.06% 7.500% 87.8% 87.4% 87.1% 3.3%10 0.49 0.68 0.70 0.7 8.60 ∗ 105 0.3204% 65.62% 7.031% 88.5% 87.7% 87.3% 3.1%11 0.47 0.60 0.75 0.9 8.81 ∗ 105 0.3281% 67.19% 6.406% 88.6% 87.9% 84.9% 5.5%12 0.47 0.60 0.75 0.9 9.01 ∗ 105 0.3357% 68.75% 6.328% 88.7% 88.0% 84.9% 5.5%13 0.48 0.66 0.70 0.9 9.22 ∗ 105 0.3433% 70.31% 5.469% 89.0% 88.7% 88.3% 2.1%14 0.48 0.66 0.70 0.9 9.42 ∗ 105 0.3510% 71.88% 5.234% 89.0% 88.6% 87.4% 3.0%15 0.48 0.65 0.70 0.9 9.63 ∗ 105 0.3586% 73.44% 5.156% 89.2% 88.8% 87.6% 2.8%16 0.48 0.65 0.75 0.9 9.83 ∗ 105 0.3662% 75.00% 5.000% 89.2% 88.5% 85.3% 5.1%17 0.49 0.67 0.70 0.7 1.00 ∗ 106 0.3738% 76.56% 4.844% 89.1% 88.4% 85.3% 5.1%18 0.49 0.67 0.75 0.8 1.02 ∗ 106 0.3815% 78.12% 4.609% 89.3% 88.6% 85.4% 5.1%19 0.49 0.67 0.75 0.9 1.04 ∗ 106 0.3891% 79.69% 4.375% 89.4% 88.8% 85.5% 4.9%20 0.47 0.66 0.75 0.9 1.06 ∗ 106 0.3967% 81.25% 4.219% 89.9% 89.4% 88.0% 2.4%21 0.47 0.64 0.70 0.9 1.09 ∗ 106 0.4044% 82.81% 4.375% 89.7% 89.0% 88.0% 2.4%22 0.46 0.62 0.75 0.9 1.11 ∗ 106 0.4120% 84.38% 3.828% 89.6% 88.4% 81.6% 8.8%23 0.46 0.62 0.75 0.8 1.13 ∗ 106 0.4196% 85.94% 3.750% 89.7% 88.4% 81.6% 8.8%24 0.47 0.64 0.70 0.9 1.15 ∗ 106 0.4272% 87.50% 3.906% 90.1% 89.4% 88.4% 2.0%25 0.47 0.66 0.75 0.9 1.17 ∗ 106 0.4349% 89.06% 3.906% 90.2% 89.5% 88.4% 2.0%26 0.48 0.64 0.65 0.9 1.19 ∗ 106 0.4425% 90.62% 3.750% 90.1% 89.4% 88.4% 2.0%27 0.48 0.64 0.65 0.9 1.21 ∗ 106 0.4501% 92.19% 3.594% 90.2% 89.5% 88.5% 1.9%28 0.48 0.64 0.65 0.9 1.23 ∗ 106 0.4578% 93.75% 3.594% 90.2% 89.5% 88.5% 1.9%29 0.48 0.64 0.65 0.9 1.25 ∗ 106 0.4654% 95.31% 3.672% 90.2% 89.5% 88.5% 1.9%30 0.48 0.64 0.65 0.9 1.27 ∗ 106 0.4730% 96.88% 3.672% 90.2% 89.5% 88.5% 1.9%31 0.47 0.64 0.70 0.9 1.29 ∗ 106 0.4807% 98.44% 3.594% 90.5% 89.8% 88.8% 1.7%32 0.47 0.64 0.70 0.9 1.31 ∗ 106 0.4883% 100.0% 3.594% 90.5% 89.8% 88.8% 1.7%33 0.47 0.60 0.75 0.9 1.33 ∗ 106 0.4959% 101.6% 3.828% 89.9% 88.8% 81.8% 8.6%34 0.49 0.63 0.65 0.5 1.35 ∗ 106 0.5035% 103.1% 4.062% 90.4% 88.8% 86.0% 4.4%35 0.49 0.63 0.65 0.5 1.37 ∗ 106 0.5112% 104.7% 4.062% 90.4% 88.8% 86.0% 4.4%36 0.49 0.63 0.65 0.5 1.39 ∗ 106 0.5188% 106.2% 4.062% 90.4% 88.8% 86.0% 4.4%37 0.49 0.63 0.65 0.5 1.41 ∗ 106 0.5264% 107.8% 4.062% 90.4% 88.8% 86.0% 4.4%38 0.49 0.63 0.65 0.5 1.43 ∗ 106 0.5341% 109.4% 4.062% 90.4% 88.8% 86.0% 4.4%39 0.46 0.70 0.70 0.9 1.45 ∗ 106 0.5417% 110.9% 3.438% 90.4% 89.7% 88.7% 1.7%40 0.46 0.70 0.70 0.9 1.47 ∗ 106 0.5493% 112.5% 3.438% 90.4% 89.7% 88.7% 1.7%41 0.46 0.70 0.70 0.9 1.50 ∗ 106 0.5569% 114.1% 3.359% 90.4% 89.7% 88.7% 1.7%42 0.46 0.70 0.70 0.9 1.52 ∗ 106 0.5646% 115.6% 3.359% 90.4% 89.7% 88.7% 1.7%43 0.46 0.70 0.70 0.9 1.54 ∗ 106 0.5722% 117.2% 3.359% 90.4% 89.7% 88.7% 1.7%44 0.46 0.70 0.70 0.9 1.56 ∗ 106 0.5798% 118.8% 3.359% 90.4% 89.7% 88.7% 1.7%45 0.46 0.70 0.70 0.9 1.58 ∗ 106 0.5875% 120.3% 3.359% 90.4% 89.7% 88.7% 1.7%46 0.46 0.70 0.70 0.9 1.60 ∗ 106 0.5951% 121.9% 3.359% 90.4% 89.7% 88.7% 1.7%47 0.46 0.70 0.70 0.9 1.62 ∗ 106 0.6027% 123.4% 3.359% 90.4% 89.7% 88.7% 1.7%48 0.46 0.70 0.70 0.9 1.64 ∗ 106 0.6104% 125.0% 3.359% 90.4% 89.7% 88.7% 1.7%* 𝟋 compared to the PSML ω256.† 𝟋 compared to the naïve binary PSML-CPCA ω256.

Table A.22: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using SMR-CPCA-M templates at T = 64 with an additional realSML-CPCA comparison and masking out most common bits.

155

Workload Ident. Verif. Recognition Performance

Approach ω256 𝟋 TP0.5 TP0.1 TP0 EER Loss (TP0)

SML-Baseline 2.68×108 − 78.1% 75.1% 72.4% 3.4% −Binary SML-CPCA (MT= 0.75) 1.31×106 0.4883% 77.5% 75.3% 73.5% 3.2% +1.1%

Table A.23: Summary of achieved TP0 with corresponding TP0.1 and TP0.5 for SMR biometric iden-tification system using all sessions of the PolyU dataset and their corresponding verification EER withapplied feature reduction approaches, ordered by TP0.

2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 4880

82

84

86

88

90

92

t

TP

0%

PSML-BaselineBf CPCA ABCBf CPCA ARCC-T SCMC-T SCM MaskingC-T SCM ARC

(a) TP0

2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 4880

82

84

86

88

90

92

t

TP

0.1

%

PSML-BaselineBf CPCA ABCBf CPCA ARCC-T SCMC-T SCM MaskingC-T SCM ARC

(b) TP0.1

Figure A.6: TP0 and TP0.1 values for relevant experiments at each t. Used abbreviations: Bf (Bloomfilter), C-T (CPCA-Tree), ABC (Additional Binary Comparison), ARC (Additional Real Comparison),SCM (SMR-CPCA-M).

156

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 5 5 0.1 4.51 ∗ 105 0.1679% 34.39% 67.4% 65.6% 58.1% 14.3%2 6 6 0.4 6.60 ∗ 105 0.2460% 50.39% 66.8% 65.8% 62.2% 10.2%3 8 6 0.3 5.24 ∗ 105 0.1952% 39.97% 69.4% 68.2% 64.9% 7.5%4 8 4 0.2 2.10 ∗ 105 0.0782% 16.02% 71.8% 68.5% 65.4% 7.0%5 8 4 0.2 2.20 ∗ 105 0.0820% 16.80% 73.7% 69.9% 66.5% 5.9%6 8 4 0.5 2.30 ∗ 105 0.0858% 17.58% 74.6% 70.7% 66.8% 5.6%7 8 4 0.5 2.41 ∗ 105 0.0896% 18.36% 75.7% 71.8% 67.8% 4.6%8 8 4 0.5 2.51 ∗ 105 0.0935% 19.14% 76.3% 72.5% 68.4% 4.0%9 7 6 0.9 7.85 ∗ 105 0.2926% 59.91% 71.3% 69.6% 65.9% 6.5%10 7 6 0.9 8.17 ∗ 105 0.3042% 62.30% 71.9% 70.1% 66.4% 6.0%11 7 6 0.9 8.48 ∗ 105 0.3158% 64.68% 72.1% 70.6% 66.8% 5.6%12 7 6 0.9 8.79 ∗ 105 0.3274% 67.06% 72.7% 71.1% 67.2% 5.2%13 7 6 0.9 9.10 ∗ 105 0.3391% 69.44% 72.9% 70.6% 67.5% 4.9%14 7 6 0.9 9.41 ∗ 105 0.3507% 71.82% 73.1% 70.7% 67.7% 4.7%15 7 6 0.9 9.73 ∗ 105 0.3623% 74.20% 73.3% 71.0% 67.8% 4.6%16 7 6 0.9 1.00 ∗ 106 0.3739% 76.58% 73.5% 71.2% 68.0% 4.4%17 7 6 0.9 1.03 ∗ 106 0.3856% 78.96% 73.8% 71.4% 68.2% 4.2%18 7 6 0.9 1.07 ∗ 106 0.3972% 81.34% 73.9% 71.6% 68.3% 4.1%19 7 6 0.9 1.10 ∗ 106 0.4088% 83.72% 74.0% 71.7% 68.4% 4.0%20 7 6 0.9 1.13 ∗ 106 0.4204% 86.10% 74.0% 71.7% 68.4% 4.0%21 7 6 0.9 1.16 ∗ 106 0.4321% 88.49% 74.2% 71.9% 68.5% 3.9%22 7 6 0.9 1.19 ∗ 106 0.4437% 90.87% 74.3% 71.9% 68.5% 3.9%23 7 6 0.9 1.22 ∗ 106 0.4553% 93.25% 74.3% 71.9% 68.6% 3.8%24 7 6 0.9 1.25 ∗ 106 0.4669% 95.63% 74.4% 72.1% 68.8% 3.7%25 7 6 0.9 1.28 ∗ 106 0.4786% 98.01% 74.5% 72.1% 68.8% 3.6%26 7 6 0.9 1.32 ∗ 106 0.4902% 100.39% 74.4% 72.1% 68.8% 3.6%27 7 6 0.9 1.35 ∗ 106 0.5018% 102.77% 74.5% 72.2% 68.9% 3.5%28 7 6 0.9 1.38 ∗ 106 0.5134% 105.15% 74.5% 72.2% 68.9% 3.5%29 7 6 0.9 1.41 ∗ 106 0.5251% 107.53% 74.5% 72.2% 68.9% 3.5%30 7 6 0.9 1.44 ∗ 106 0.5367% 109.91% 74.6% 72.3% 68.9% 3.5%31 7 6 0.9 1.47 ∗ 106 0.5483% 112.30% 74.7% 72.3% 69.0% 3.4%32 7 6 0.9 1.50 ∗ 106 0.5599% 114.68% 74.8% 72.4% 69.0% 3.4%33 7 6 0.9 1.53 ∗ 106 0.5716% 117.06% 74.8% 72.4% 69.1% 3.3%34 7 6 0.9 1.57 ∗ 106 0.5832% 119.44% 74.8% 72.4% 69.1% 3.3%35 7 6 0.9 1.60 ∗ 106 0.5948% 121.82% 74.8% 72.4% 69.1% 3.3%36 7 6 0.9 1.63 ∗ 106 0.6064% 124.20% 74.9% 71.6% 69.1% 3.3%37 7 6 0.9 1.66 ∗ 106 0.6181% 126.58% 74.9% 71.6% 69.1% 3.3%38 7 6 0.9 1.69 ∗ 106 0.6297% 128.96% 74.9% 71.6% 69.1% 3.3%39 7 6 0.9 1.72 ∗ 106 0.6413% 131.34% 74.9% 71.7% 69.2% 3.2%40 7 6 0.9 1.75 ∗ 106 0.6529% 133.72% 74.9% 71.7% 69.2% 3.2%41 7 6 0.9 1.78 ∗ 106 0.6646% 136.10% 74.9% 71.7% 69.2% 3.2%42 7 6 0.9 1.82 ∗ 106 0.6762% 138.49% 74.9% 71.7% 69.2% 3.2%43 7 6 0.9 1.85 ∗ 106 0.6878% 140.87% 74.9% 71.7% 69.2% 3.2%44 7 6 0.9 1.88 ∗ 106 0.6995% 143.25% 75.0% 71.7% 69.2% 3.2%45 7 6 0.9 1.91 ∗ 106 0.7111% 145.63% 75.0% 71.7% 69.2% 3.2%46 7 6 0.9 1.94 ∗ 106 0.7227% 148.01% 75.0% 71.7% 69.2% 3.2%47 7 6 0.9 1.97 ∗ 106 0.7343% 150.39% 75.0% 71.7% 69.2% 3.2%48 7 6 0.9 2.00 ∗ 106 0.7460% 152.77% 75.0% 71.7% 69.2% 3.2%* 𝟋 compared to the SML ω256.† 𝟋 compared to the naïve binary SML-CPCA ω256.

Table A.24: Best achieved binary SML-CPCA Bloom filter TP0 with corresponding TP0.1 and TP0.5,with a final real-valued SML-CPCA comparison and tree pre-selection of t = 1, . . . , 48 trees at T = 64using both sessions of the PolyU dataset.

157

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.48 0.64 0.75 3.48 ∗ 105 0.1297% 26.56% 37.18% 59.5% 58.5% 56.5% 15.9%2 0.48 0.64 0.75 3.69 ∗ 105 0.1373% 28.12% 26.63% 67.5% 66.5% 63.6% 8.8%3 0.45 0.60 0.80 3.89 ∗ 105 0.1450% 29.69% 23.40% 69.9% 68.4% 64.7% 7.7%4 0.45 0.60 0.80 4.10 ∗ 105 0.1526% 31.25% 20.38% 71.9% 70.3% 66.4% 6.0%5 0.45 0.60 0.80 4.30 ∗ 105 0.1602% 32.81% 18.00% 72.9% 71.8% 67.8% 4.6%6 0.45 0.60 0.80 4.51 ∗ 105 0.1678% 34.38% 16.34% 74.1% 72.9% 68.8% 3.6%7 0.45 0.60 0.80 4.71 ∗ 105 0.1755% 35.94% 15.20% 74.7% 73.4% 69.2% 3.2%8 0.45 0.60 0.80 4.92 ∗ 105 0.1831% 37.50% 13.60% 75.3% 74.0% 69.7% 2.7%9 0.45 0.60 0.80 5.12 ∗ 105 0.1907% 39.06% 13.03% 75.8% 74.4% 70.0% 2.4%10 0.45 0.60 0.80 5.32 ∗ 105 0.1984% 40.62% 12.50% 76.2% 73.1% 70.2% 2.2%11 0.45 0.60 0.80 5.53 ∗ 105 0.2060% 42.19% 11.75% 76.7% 74.0% 70.6% 1.8%12 0.45 0.60 0.80 5.73 ∗ 105 0.2136% 43.75% 11.12% 77.0% 74.3% 70.8% 1.6%13 0.45 0.60 0.80 5.94 ∗ 105 0.2213% 45.31% 10.83% 77.1% 74.3% 70.8% 1.6%14 0.45 0.60 0.80 6.14 ∗ 105 0.2289% 46.88% 10.58% 77.2% 74.5% 70.9% 1.5%15 0.45 0.60 0.80 6.35 ∗ 105 0.2365% 48.44% 10.44% 77.3% 74.5% 70.9% 1.5%16 0.45 0.60 0.80 6.55 ∗ 105 0.2441% 50.00% 10.30% 77.3% 74.5% 70.9% 1.5%17 0.45 0.60 0.80 6.76 ∗ 105 0.2518% 51.56% 10.05% 77.2% 74.6% 71.0% 1.4%18 0.45 0.60 0.80 6.96 ∗ 105 0.2594% 53.12% 9.80% 77.3% 74.6% 71.0% 1.4%19 0.45 0.60 0.80 7.17 ∗ 105 0.2670% 54.69% 9.59% 77.3% 74.7% 71.0% 1.4%20 0.45 0.60 0.80 7.37 ∗ 105 0.2747% 56.25% 9.38% 77.4% 74.8% 71.1% 1.3%21 0.45 0.60 0.80 7.58 ∗ 105 0.2823% 57.81% 9.38% 77.4% 74.8% 71.1% 1.3%22 0.45 0.60 0.80 7.78 ∗ 105 0.2899% 59.38% 9.41% 77.4% 74.8% 71.1% 1.3%23 0.45 0.60 0.80 7.99 ∗ 105 0.2975% 60.94% 9.30% 77.5% 74.8% 71.1% 1.3%24 0.45 0.60 0.80 8.19 ∗ 105 0.3052% 62.50% 9.23% 77.5% 74.9% 71.1% 1.3%25 0.45 0.60 0.80 8.40 ∗ 105 0.3128% 64.06% 9.09% 77.5% 74.9% 71.1% 1.3%26 0.45 0.60 0.80 8.60 ∗ 105 0.3204% 65.62% 9.06% 77.5% 74.9% 71.1% 1.3%27 0.45 0.60 0.80 8.81 ∗ 105 0.3281% 67.19% 8.88% 77.6% 74.9% 71.1% 1.3%28 0.45 0.60 0.80 9.01 ∗ 105 0.3357% 68.75% 8.74% 77.6% 74.9% 71.2% 1.2%29 0.45 0.60 0.80 9.22 ∗ 105 0.3433% 70.31% 8.66% 77.6% 74.9% 71.2% 1.2%30 0.45 0.60 0.80 9.42 ∗ 105 0.3510% 71.88% 8.59% 77.6% 75.0% 71.2% 1.2%31 0.45 0.60 0.80 9.63 ∗ 105 0.3586% 73.44% 8.49% 77.7% 75.0% 71.2% 1.2%32 0.45 0.60 0.80 9.83 ∗ 105 0.3662% 75.00% 8.31% 77.7% 75.0% 71.2% 1.2%33 0.45 0.60 0.80 1.00 ∗ 106 0.3738% 76.56% 8.27% 77.7% 75.0% 71.2% 1.2%34 0.45 0.60 0.80 1.02 ∗ 106 0.3815% 78.12% 8.24% 77.7% 75.0% 71.2% 1.2%35 0.45 0.60 0.80 1.04 ∗ 106 0.3891% 79.69% 8.24% 77.7% 75.0% 71.2% 1.2%36 0.45 0.60 0.80 1.06 ∗ 106 0.3967% 81.25% 8.17% 77.7% 75.1% 71.3% 1.1%37 0.45 0.60 0.80 1.09 ∗ 106 0.4044% 82.81% 8.17% 77.7% 75.1% 71.3% 1.1%38 0.45 0.60 0.80 1.11 ∗ 106 0.4120% 84.38% 8.13% 77.7% 75.1% 71.3% 1.1%39 0.45 0.60 0.80 1.13 ∗ 106 0.4196% 85.94% 8.13% 77.7% 75.1% 71.3% 1.1%40 0.45 0.60 0.80 1.15 ∗ 106 0.4272% 87.50% 8.13% 77.7% 75.1% 71.3% 1.1%41 0.45 0.60 0.80 1.17 ∗ 106 0.4349% 89.06% 8.10% 77.7% 75.1% 71.3% 1.1%42 0.45 0.60 0.80 1.19 ∗ 106 0.4425% 90.62% 8.03% 77.8% 75.1% 71.3% 1.1%43 0.45 0.60 0.80 1.21 ∗ 106 0.4501% 92.19% 7.99% 77.8% 75.1% 71.3% 1.1%44 0.45 0.60 0.80 1.23 ∗ 106 0.4578% 93.75% 7.99% 77.8% 75.1% 71.3% 1.1%45 0.45 0.60 0.80 1.25 ∗ 106 0.4654% 95.31% 7.95% 77.8% 75.1% 71.3% 1.1%46 0.45 0.60 0.80 1.27 ∗ 106 0.4730% 96.88% 7.92% 77.8% 75.1% 71.3% 1.1%47 0.45 0.60 0.80 1.29 ∗ 106 0.4807% 98.44% 7.92% 77.8% 75.1% 71.3% 1.1%48 0.45 0.60 0.80 1.31 ∗ 106 0.4883% 100.00% 7.92% 77.8% 75.1% 71.3% 1.1%* 𝟋 compared to the SML ω256.† 𝟋 compared to the naïve binary SML-CPCA ω256.

Table A.25: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using SMR-CPCA-M templates at T = 64 with both sessions of thePolyU dataset.

158

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.48 0.64 0.75 6.76 ∗ 105 0.2518% 51.56% 37.18% 59.7% 58.3% 57.0% 15.4%2 0.48 0.64 0.75 6.96 ∗ 105 0.2594% 53.12% 26.63% 68.1% 66.3% 64.7% 7.7%3 0.45 0.60 0.80 7.17 ∗ 105 0.2670% 54.69% 23.40% 69.0% 67.2% 65.5% 6.9%4 0.45 0.60 0.80 7.37 ∗ 105 0.2747% 56.25% 20.38% 71.0% 68.8% 67.2% 5.2%5 0.45 0.60 0.80 7.58 ∗ 105 0.2823% 57.81% 18.00% 72.8% 70.5% 68.7% 3.7%6 0.45 0.60 0.80 7.78 ∗ 105 0.2899% 59.38% 16.34% 73.7% 71.4% 69.5% 2.9%7 0.45 0.60 0.80 7.99 ∗ 105 0.2975% 60.94% 15.20% 74.2% 72.2% 69.9% 2.5%8 0.45 0.60 0.80 8.19 ∗ 105 0.3052% 62.50% 13.60% 74.8% 72.6% 70.3% 2.1%9 0.45 0.60 0.80 8.40 ∗ 105 0.3128% 64.06% 13.03% 75.2% 72.9% 70.5% 1.9%10 0.45 0.60 0.83 8.60 ∗ 105 0.3204% 65.62% 12.50% 75.5% 73.3% 70.9% 1.5%11 0.45 0.60 0.80 8.81 ∗ 105 0.3281% 67.19% 11.75% 75.9% 73.5% 71.1% 1.3%12 0.45 0.60 0.80 9.01 ∗ 105 0.3357% 68.75% 11.12% 76.2% 73.9% 71.3% 1.1%13 0.45 0.60 0.80 9.22 ∗ 105 0.3433% 70.31% 10.83% 76.3% 73.9% 71.4% 1.0%14 0.45 0.60 0.80 9.42 ∗ 105 0.3510% 71.88% 10.58% 76.5% 74.1% 71.5% 0.9%15 0.45 0.60 0.80 9.63 ∗ 105 0.3586% 73.44% 10.44% 76.6% 74.1% 71.5% 0.9%16 0.45 0.60 0.80 9.83 ∗ 105 0.3662% 75.00% 10.30% 76.4% 74.1% 71.5% 0.9%17 0.45 0.60 0.80 1.00 ∗ 106 0.3738% 76.56% 10.05% 76.5% 74.2% 71.6% 0.8%18 0.45 0.60 0.80 1.02 ∗ 106 0.3815% 78.12% 9.80% 76.6% 74.2% 71.6% 0.8%19 0.45 0.60 0.80 1.04 ∗ 106 0.3891% 79.69% 9.59% 76.7% 74.3% 71.6% 0.8%20 0.45 0.60 0.80 1.06 ∗ 106 0.3967% 81.25% 9.38% 76.7% 74.3% 71.6% 0.8%21 0.45 0.60 0.80 1.09 ∗ 106 0.4044% 82.81% 9.38% 76.7% 74.3% 71.6% 0.8%22 0.45 0.60 0.80 1.11 ∗ 106 0.4120% 84.38% 9.41% 76.7% 74.3% 71.6% 0.8%23 0.45 0.60 0.80 1.13 ∗ 106 0.4196% 85.94% 9.30% 76.8% 74.4% 71.7% 0.7%24 0.45 0.60 0.80 1.15 ∗ 106 0.4272% 87.50% 9.23% 76.8% 74.4% 71.7% 0.7%25 0.45 0.60 0.80 1.17 ∗ 106 0.4349% 89.06% 9.09% 76.8% 74.4% 71.7% 0.7%26 0.45 0.60 0.80 1.19 ∗ 106 0.4425% 90.62% 9.06% 76.8% 74.4% 71.7% 0.7%27 0.45 0.60 0.80 1.21 ∗ 106 0.4501% 92.19% 8.88% 76.9% 74.5% 71.7% 0.7%28 0.45 0.60 0.80 1.23 ∗ 106 0.4578% 93.75% 8.74% 77.0% 74.5% 71.8% 0.6%29 0.45 0.60 0.80 1.25 ∗ 106 0.4654% 95.31% 8.66% 77.0% 74.5% 71.8% 0.6%30 0.45 0.60 0.80 1.27 ∗ 106 0.4730% 96.88% 8.59% 77.0% 74.6% 71.8% 0.6%31 0.45 0.60 0.80 1.29 ∗ 106 0.4807% 98.44% 8.49% 77.1% 74.6% 71.8% 0.6%32 0.45 0.60 0.80 1.31 ∗ 106 0.4883% 100.00% 8.31% 77.1% 74.6% 71.9% 0.5%33 0.45 0.60 0.80 1.33 ∗ 106 0.4959% 101.56% 8.27% 77.1% 74.6% 71.9% 0.5%34 0.45 0.60 0.80 1.35 ∗ 106 0.5035% 103.12% 8.24% 77.2% 74.7% 71.9% 0.5%35 0.45 0.60 0.80 1.37 ∗ 106 0.5112% 104.69% 8.24% 77.2% 74.7% 71.9% 0.5%36 0.45 0.60 0.80 1.39 ∗ 106 0.5188% 106.25% 8.17% 77.2% 74.7% 71.9% 0.5%37 0.45 0.60 0.80 1.41 ∗ 106 0.5264% 107.81% 8.17% 77.2% 74.7% 71.9% 0.5%38 0.45 0.60 0.80 1.43 ∗ 106 0.5341% 109.38% 8.13% 77.2% 74.7% 71.9% 0.5%39 0.45 0.60 0.80 1.45 ∗ 106 0.5417% 110.94% 8.13% 77.2% 74.7% 71.9% 0.5%40 0.45 0.60 0.80 1.47 ∗ 106 0.5493% 112.50% 8.13% 77.2% 74.7% 71.9% 0.5%41 0.45 0.60 0.80 1.50 ∗ 106 0.5569% 114.06% 8.10% 77.2% 74.7% 71.9% 0.5%42 0.45 0.60 0.80 1.52 ∗ 106 0.5646% 115.62% 8.03% 77.2% 74.7% 71.9% 0.5%43 0.45 0.60 0.80 1.54 ∗ 106 0.5722% 117.19% 7.99% 77.3% 74.7% 71.9% 0.5%44 0.45 0.60 0.80 1.56 ∗ 106 0.5798% 118.75% 7.99% 77.3% 74.7% 71.9% 0.5%45 0.45 0.60 0.80 1.58 ∗ 106 0.5875% 120.31% 7.95% 77.3% 74.7% 71.9% 0.5%46 0.45 0.60 0.80 1.60 ∗ 106 0.5951% 121.88% 7.92% 77.3% 74.7% 71.9% 0.5%47 0.45 0.60 0.80 1.62 ∗ 106 0.6027% 123.44% 7.92% 77.3% 74.7% 71.9% 0.5%48 0.45 0.60 0.80 1.64 ∗ 106 0.6104% 125.00% 7.92% 77.3% 74.7% 71.9% 0.5%* 𝟋 compared to the SML ω256.† 𝟋 compared to the naïve binary SML-CPCA ω256.

Table A.26: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using SMR-CPCA-M templates at T = 64 with an additional realSML-CPCA comparison employing both sessions of the PolyU dataset.

159

SMR Workload Recognition Performance

t λmax λCP CAmax MT Bit Threshold ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.46 0.60 0.75 0.7 3.48 ∗ 105 0.1297% 26.56% 37.93% 59.3% 57.8% 56.9% 15.5%2 0.45 0.60 0.75 0.7 3.69 ∗ 105 0.1373% 28.12% 27.81% 66.6% 65.1% 62.6% 9.8%3 0.45 0.60 0.80 0.7 3.89 ∗ 105 0.1450% 29.69% 23.54% 70.0% 68.6% 65.2% 7.2%4 0.45 0.60 0.80 0.7 4.10 ∗ 105 0.1526% 31.25% 20.38% 72.1% 70.1% 67.0% 5.4%5 0.45 0.60 0.80 0.7 4.30 ∗ 105 0.1602% 32.81% 18.08% 73.5% 71.0% 68.4% 4.0%6 0.45 0.60 0.80 0.7 4.51 ∗ 105 0.1678% 34.38% 16.34% 74.9% 72.2% 69.5% 2.9%7 0.45 0.60 0.80 0.7 4.71 ∗ 105 0.1755% 35.94% 15.27% 75.5% 72.6% 69.9% 2.5%8 0.45 0.60 0.80 0.7 4.92 ∗ 105 0.1831% 37.50% 13.67% 76.1% 73.2% 70.4% 2.0%9 0.45 0.60 0.80 0.7 5.12 ∗ 105 0.1907% 39.06% 13.21% 76.5% 73.5% 70.6% 1.8%10 0.45 0.60 0.80 0.7 5.32 ∗ 105 0.1984% 40.62% 12.50% 77.0% 73.8% 70.9% 1.5%11 0.45 0.60 0.80 0.7 5.53 ∗ 105 0.2060% 42.19% 12.00% 77.1% 73.8% 71.2% 1.2%12 0.45 0.60 0.80 0.7 5.73 ∗ 105 0.2136% 43.75% 11.40% 77.4% 74.1% 71.4% 1.0%13 0.45 0.60 0.80 0.7 5.94 ∗ 105 0.2213% 45.31% 11.08% 77.5% 74.2% 71.5% 0.9%14 0.45 0.60 0.80 0.7 6.14 ∗ 105 0.2289% 46.88% 10.94% 77.7% 74.3% 71.6% 0.8%15 0.45 0.60 0.80 0.7 6.35 ∗ 105 0.2365% 48.44% 10.51% 77.8% 74.4% 71.6% 0.8%16 0.45 0.60 0.80 0.7 6.55 ∗ 105 0.2441% 50.00% 10.48% 77.8% 74.4% 71.6% 0.8%17 0.45 0.60 0.80 0.7 6.76 ∗ 105 0.2518% 51.56% 10.30% 77.3% 74.5% 71.7% 0.7%18 0.45 0.60 0.80 0.7 6.96 ∗ 105 0.2594% 53.12% 9.98% 77.4% 74.5% 71.7% 0.7%19 0.45 0.60 0.80 0.7 7.17 ∗ 105 0.2670% 54.69% 9.73% 77.5% 74.6% 71.8% 0.6%20 0.45 0.60 0.80 0.7 7.37 ∗ 105 0.2747% 56.25% 9.62% 77.5% 74.6% 71.8% 0.6%21 0.45 0.60 0.80 0.7 7.58 ∗ 105 0.2823% 57.81% 9.48% 77.6% 74.7% 71.9% 0.5%22 0.45 0.60 0.80 0.7 7.78 ∗ 105 0.2899% 59.38% 9.48% 77.6% 74.7% 71.9% 0.5%23 0.45 0.60 0.80 0.7 7.99 ∗ 105 0.2975% 60.94% 9.34% 77.6% 74.7% 71.9% 0.5%24 0.45 0.60 0.80 0.7 8.19 ∗ 105 0.3052% 62.50% 9.27% 77.6% 74.7% 71.9% 0.5%25 0.45 0.60 0.80 0.7 8.40 ∗ 105 0.3128% 64.06% 9.13% 77.6% 74.8% 71.9% 0.5%26 0.45 0.60 0.80 0.7 8.60 ∗ 105 0.3204% 65.62% 9.09% 77.6% 74.8% 71.9% 0.5%27 0.45 0.60 0.80 0.7 8.81 ∗ 105 0.3281% 67.19% 8.95% 77.7% 74.8% 71.9% 0.5%28 0.45 0.60 0.80 0.7 9.01 ∗ 105 0.3357% 68.75% 8.81% 77.7% 74.8% 72.0% 0.4%29 0.45 0.60 0.80 0.7 9.22 ∗ 105 0.3433% 70.31% 8.77% 77.7% 74.8% 72.0% 0.4%30 0.45 0.60 0.80 0.7 9.42 ∗ 105 0.3510% 71.88% 8.63% 77.7% 74.9% 72.0% 0.4%31 0.45 0.60 0.80 0.7 9.63 ∗ 105 0.3586% 73.44% 8.56% 77.8% 74.9% 72.1% 0.3%32 0.45 0.60 0.80 0.7 9.83 ∗ 105 0.3662% 75.00% 8.38% 77.8% 74.9% 72.1% 0.3%33 0.45 0.60 0.80 0.7 1.00 ∗ 106 0.3738% 76.56% 8.35% 77.8% 74.9% 72.1% 0.3%34 0.45 0.60 0.80 0.7 1.02 ∗ 106 0.3815% 78.12% 8.31% 77.8% 74.9% 72.1% 0.3%35 0.45 0.60 0.80 0.7 1.04 ∗ 106 0.3891% 79.69% 8.31% 77.8% 74.9% 72.1% 0.3%36 0.45 0.60 0.80 0.7 1.06 ∗ 106 0.3967% 81.25% 8.24% 77.9% 75.0% 72.1% 0.3%37 0.45 0.60 0.80 0.7 1.09 ∗ 106 0.4044% 82.81% 8.24% 77.9% 75.0% 72.1% 0.3%38 0.45 0.60 0.80 0.7 1.11 ∗ 106 0.4120% 84.38% 8.20% 77.9% 75.0% 72.1% 0.3%39 0.45 0.60 0.80 0.7 1.13 ∗ 106 0.4196% 85.94% 8.17% 77.9% 75.0% 72.1% 0.3%40 0.45 0.60 0.80 0.7 1.15 ∗ 106 0.4272% 87.50% 8.17% 77.9% 75.0% 72.1% 0.3%41 0.45 0.60 0.80 0.7 1.17 ∗ 106 0.4349% 89.06% 8.13% 77.9% 75.0% 72.1% 0.3%42 0.45 0.60 0.80 0.7 1.19 ∗ 106 0.4425% 90.62% 8.10% 77.9% 75.0% 72.1% 0.3%43 0.45 0.60 0.80 0.7 1.21 ∗ 106 0.4501% 92.19% 8.03% 77.9% 75.0% 72.1% 0.3%44 0.45 0.60 0.80 0.7 1.23 ∗ 106 0.4578% 93.75% 7.99% 77.9% 75.0% 72.1% 0.3%45 0.45 0.60 0.80 0.7 1.25 ∗ 106 0.4654% 95.31% 7.99% 77.9% 75.0% 72.1% 0.3%46 0.45 0.60 0.80 0.7 1.27 ∗ 106 0.4730% 96.88% 7.95% 77.9% 75.0% 72.1% 0.3%47 0.45 0.60 0.80 0.7 1.29 ∗ 106 0.4807% 98.44% 7.95% 77.9% 75.0% 72.1% 0.3%48 0.45 0.60 0.80 0.7 1.31 ∗ 106 0.4883% 100.00% 7.95% 77.9% 75.0% 72.1% 0.3%* 𝟋 compared to the SML ω256.† 𝟋 compared to the naïve binary SML-CPCA ω256.

Table A.27: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using SMR-CPCA-M templates at T = 64 with an additional binarySML-CPCA comparison and masking out most common set bits employing both sessions of the PolyUdataset.

160

Workload Ident. Verif. Recognition Performance

Approach ω256 𝟋 TP0.5 TP0.1 TP0 EER Loss (TP0)

SML-Baseline 2.68×108 − 89.9% 88.4% 88.1% 2.3% −Binary SML-CPCA (MT= 0.75) 1.31×106 0.4883% 88.3% 86.3% 86% 2.4% 2.1%

Table A.28: Summary of achieved TP0 with corresponding TP0.1 and TP0.5 for the SMR biometricidentification system using session 2 of the PolyU dataset and their corresponding verification EERwith applied feature reduction approaches, ordered by TP0.

161

Bloom filter SMR Workload Recognition Performance

t W H MT ω256 𝟋B* 𝟋R

† TP0.5 TP0.1 TP0TP0 loss to

PSML-Baseline

1 5 5 0.1 4.51 ∗ 105 0.1679% 34.39% 81.2% 80.9% 80.5% 7.7%2 4 4 0.2 3.74 ∗ 105 0.1392% 28.52% 84.8% 84.4% 83.8% 4.3%3 4 4 0.2 3.94 ∗ 105 0.1469% 30.08% 86.2% 85.4% 85.2% 2.9%4 4 4 0.5 4.15 ∗ 105 0.1545% 31.64% 87.1% 86.5% 83.8% 4.3%5 4 5 0.4 6.93 ∗ 105 0.2583% 52.89% 87.7% 87.0% 84.3% 3.8%6 7 4 0.1 2.63 ∗ 105 0.0978% 20.03% 86.2% 85.5% 85.4% 2.7%7 7 4 0.1 2.74 ∗ 105 0.1022% 20.93% 86.6% 85.9% 85.9% 2.3%8 7 4 0.1 2.86 ∗ 105 0.1065% 21.82% 86.9% 86.2% 86.1% 2.0%9 7 4 0.1 2.98 ∗ 105 0.1109% 22.71% 87.4% 86.6% 86.5% 1.6%10 7 4 0.1 3.09 ∗ 105 0.1153% 23.60% 87.7% 86.8% 86.7% 1.4%11 7 5 0.6 5.11 ∗ 105 0.1902% 38.96% 87.2% 86.9% 86.4% 1.7%12 8 4 0.2 2.92 ∗ 105 0.1087% 22.27% 87.6% 86.9% 86.6% 1.5%13 7 5 0.6 5.48 ∗ 105 0.2042% 41.82% 87.7% 87.3% 86.8% 1.3%14 7 5 0.6 5.67 ∗ 105 0.2112% 43.25% 87.7% 87.4% 87.0% 1.2%15 7 5 0.6 5.86 ∗ 105 0.2181% 44.68% 88.0% 87.6% 87.1% 1.0%16 7 5 0.6 6.04 ∗ 105 0.2251% 46.10% 88.0% 87.7% 87.3% 0.9%17 7 5 0.6 6.23 ∗ 105 0.2321% 47.53% 88.3% 88.0% 87.5% 0.6%18 7 5 0.6 6.42 ∗ 105 0.2391% 48.96% 88.3% 88.0% 87.5% 0.6%19 7 5 0.6 6.60 ∗ 105 0.2460% 50.39% 88.4% 88.0% 87.6% 0.5%20 7 5 0.6 6.79 ∗ 105 0.2530% 51.82% 88.4% 88.0% 87.6% 0.5%21 7 5 0.6 6.98 ∗ 105 0.2600% 53.25% 88.4% 88.1% 87.6% 0.5%22 7 5 0.6 7.17 ∗ 105 0.2670% 54.68% 88.5% 88.2% 87.7% 0.4%23 7 5 0.6 7.35 ∗ 105 0.2739% 56.10% 88.4% 88.1% 87.6% 0.5%24 7 5 0.6 7.54 ∗ 105 0.2809% 57.53% 88.4% 88.1% 87.6% 0.5%25 7 5 0.6 7.73 ∗ 105 0.2879% 58.96% 88.6% 88.3% 87.7% 0.4%26 7 5 0.6 7.92 ∗ 105 0.2949% 60.39% 88.6% 88.3% 87.7% 0.4%27 7 5 0.6 8.10 ∗ 105 0.3019% 61.82% 88.6% 88.3% 87.7% 0.4%28 7 5 0.6 8.29 ∗ 105 0.3088% 63.25% 88.6% 88.3% 87.7% 0.4%29 7 5 0.6 8.48 ∗ 105 0.3158% 64.68% 88.5% 88.2% 87.7% 0.4%30 7 5 0.6 8.66 ∗ 105 0.3228% 66.10% 88.7% 88.4% 87.7% 0.4%31 7 5 0.6 8.85 ∗ 105 0.3298% 67.53% 88.7% 88.4% 87.7% 0.4%32 7 5 0.6 9.04 ∗ 105 0.3367% 68.96% 88.7% 88.4% 87.7% 0.4%33 7 5 0.6 9.23 ∗ 105 0.3437% 70.39% 88.7% 88.4% 87.7% 0.4%34 7 5 0.6 9.41 ∗ 105 0.3507% 71.82% 88.7% 88.4% 87.7% 0.4%35 7 5 0.6 9.60 ∗ 105 0.3577% 73.25% 88.6% 88.3% 87.7% 0.4%36 7 5 0.6 9.79 ∗ 105 0.3646% 74.68% 88.6% 88.3% 87.7% 0.4%37 7 5 0.6 9.98 ∗ 105 0.3716% 76.10% 88.6% 88.3% 87.7% 0.4%38 7 5 0.6 1.02 ∗ 106 0.3786% 77.53% 88.6% 88.3% 87.7% 0.4%39 7 5 0.6 1.03 ∗ 106 0.3856% 78.96% 88.6% 88.3% 87.7% 0.4%40 7 5 0.6 1.05 ∗ 106 0.3925% 80.39% 88.6% 88.3% 87.7% 0.4%41 7 5 0.6 1.07 ∗ 106 0.3995% 81.82% 88.6% 88.3% 87.7% 0.4%42 7 5 0.6 1.09 ∗ 106 0.4065% 83.25% 88.6% 88.3% 87.7% 0.4%43 7 5 0.6 1.11 ∗ 106 0.4135% 84.68% 88.6% 88.3% 87.7% 0.4%44 7 5 0.6 1.13 ∗ 106 0.4204% 86.10% 88.6% 88.3% 87.7% 0.4%45 7 5 0.6 1.15 ∗ 106 0.4274% 87.53% 88.6% 88.3% 87.7% 0.4%46 7 5 0.6 1.17 ∗ 106 0.4344% 88.96% 88.6% 88.3% 87.7% 0.4%47 7 5 0.6 1.18 ∗ 106 0.4414% 90.39% 88.6% 88.3% 87.7% 0.4%48 7 5 0.6 1.20 ∗ 106 0.4483% 91.82% 88.6% 88.3% 87.7% 0.4%* 𝟋 compared to the SML ω256.† 𝟋 compared to the naïve binary SML-CPCA ω256.

Table A.29: Best achieved binary SML-CPCA Bloom filter TP0 with corresponding TP0.1 and TP0.5,with a final real-valued SML-CPCA comparison and tree pre-selection of t = 1, . . . , 48 trees at T = 64using for session 2 of the PolyU dataset.

162

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.50 0.62 0.70 3.48 ∗ 105 0.1297% 26.56% 27.34% 71.3% 71.2% 70.9% 17.3%2 0.52 0.62 0.70 3.69 ∗ 105 0.1373% 28.12% 17.34% 80.4% 80.0% 79.5% 8.6%3 0.51 0.61 0.70 3.89 ∗ 105 0.1450% 29.69% 13.44% 84.1% 83.3% 82.7% 5.5%4 0.52 0.64 0.75 4.10 ∗ 105 0.1526% 31.25% 11.88% 85.4% 84.8% 84.5% 3.7%5 0.52 0.62 0.70 4.30 ∗ 105 0.1602% 32.81% 10.62% 86.4% 85.5% 85.3% 2.8%6 0.51 0.61 0.75 4.51 ∗ 105 0.1678% 34.38% 9.06% 86.7% 86.5% 85.6% 2.5%7 0.51 0.61 0.75 4.71 ∗ 105 0.1755% 35.94% 7.97% 87.4% 87.2% 86.3% 1.8%8 0.51 0.61 0.75 4.92 ∗ 105 0.1831% 37.50% 7.81% 87.5% 87.3% 86.4% 1.7%9 0.50 0.60 0.70 5.12 ∗ 105 0.1907% 39.06% 7.50% 87.6% 87.2% 86.4% 1.7%10 0.43 0.61 0.75 5.32 ∗ 105 0.1984% 40.62% 5.78% 87.7% 86.7% 86.4% 1.7%11 0.43 0.61 0.75 5.53 ∗ 105 0.2060% 42.19% 5.47% 87.8% 86.9% 86.6% 1.6%12 0.43 0.61 0.75 5.73 ∗ 105 0.2136% 43.75% 5.00% 87.8% 87.0% 86.7% 1.4%13 0.43 0.61 0.75 5.94 ∗ 105 0.2213% 45.31% 4.92% 87.9% 87.0% 86.7% 1.4%14 0.45 0.64 0.70 6.14 ∗ 105 0.2289% 46.88% 4.53% 88.7% 87.5% 86.7% 1.4%15 0.44 0.63 0.70 6.35 ∗ 105 0.2365% 48.44% 4.45% 88.9% 88.1% 87.0% 1.2%16 0.43 0.61 0.75 6.55 ∗ 105 0.2441% 50.00% 4.38% 88.4% 87.4% 87.1% 1.0%17 0.43 0.61 0.75 6.76 ∗ 105 0.2518% 51.56% 4.22% 88.4% 87.4% 87.1% 1.0%18 0.43 0.61 0.75 6.96 ∗ 105 0.2594% 53.12% 4.14% 88.4% 87.5% 87.2% 0.9%19 0.43 0.61 0.75 7.17 ∗ 105 0.2670% 54.69% 3.91% 88.6% 87.7% 87.3% 0.8%20 0.43 0.61 0.75 7.37 ∗ 105 0.2747% 56.25% 3.91% 88.6% 87.7% 87.3% 0.8%21 0.43 0.61 0.75 7.58 ∗ 105 0.2823% 57.81% 3.91% 88.6% 87.7% 87.3% 0.8%22 0.43 0.61 0.75 7.78 ∗ 105 0.2899% 59.38% 3.83% 88.7% 87.7% 87.4% 0.7%23 0.43 0.61 0.75 7.99 ∗ 105 0.2975% 60.94% 3.83% 88.7% 87.7% 87.4% 0.7%24 0.43 0.61 0.75 8.19 ∗ 105 0.3052% 62.50% 3.83% 88.7% 87.7% 87.4% 0.7%25 0.43 0.61 0.75 8.40 ∗ 105 0.3128% 64.06% 3.91% 88.7% 87.7% 87.4% 0.7%26 0.43 0.61 0.75 8.60 ∗ 105 0.3204% 65.62% 3.83% 88.7% 87.7% 87.4% 0.7%27 0.43 0.61 0.75 8.81 ∗ 105 0.3281% 67.19% 3.83% 88.7% 87.7% 87.4% 0.7%28 0.43 0.61 0.75 9.01 ∗ 105 0.3357% 68.75% 3.83% 88.7% 87.7% 87.4% 0.7%29 0.44 0.62 0.75 9.22 ∗ 105 0.3433% 70.31% 4.06% 89.1% 87.6% 87.5% 0.6%30 0.44 0.62 0.75 9.42 ∗ 105 0.3510% 71.88% 3.98% 89.1% 87.6% 87.5% 0.6%31 0.44 0.62 0.75 9.63 ∗ 105 0.3586% 73.44% 3.98% 88.8% 87.6% 87.5% 0.6%32 0.44 0.62 0.75 9.83 ∗ 105 0.3662% 75.00% 3.98% 88.8% 87.6% 87.5% 0.6%33 0.44 0.62 0.75 1.00 ∗ 106 0.3738% 76.56% 3.98% 88.8% 87.6% 87.5% 0.6%34 0.44 0.62 0.75 1.02 ∗ 106 0.3815% 78.12% 3.91% 88.8% 87.6% 87.5% 0.6%35 0.44 0.62 0.75 1.04 ∗ 106 0.3891% 79.69% 3.91% 88.8% 87.6% 87.5% 0.6%36 0.44 0.62 0.75 1.06 ∗ 106 0.3967% 81.25% 3.91% 88.8% 87.6% 87.5% 0.6%37 0.44 0.62 0.75 1.09 ∗ 106 0.4044% 82.81% 3.91% 88.8% 87.6% 87.5% 0.6%38 0.44 0.62 0.75 1.11 ∗ 106 0.4120% 84.38% 3.91% 88.8% 87.6% 87.5% 0.6%39 0.44 0.62 0.75 1.13 ∗ 106 0.4196% 85.94% 3.91% 88.8% 87.6% 87.5% 0.6%40 0.44 0.62 0.75 1.15 ∗ 106 0.4272% 87.50% 3.91% 88.8% 87.6% 87.5% 0.6%41 0.44 0.62 0.75 1.17 ∗ 106 0.4349% 89.06% 3.91% 88.8% 87.6% 87.5% 0.6%42 0.54 0.64 0.75 1.19 ∗ 106 0.4425% 90.62% 4.06% 88.5% 87.7% 87.5% 0.6%43 0.54 0.64 0.75 1.21 ∗ 106 0.4501% 92.19% 4.14% 88.5% 87.7% 87.5% 0.6%44 0.54 0.64 0.75 1.23 ∗ 106 0.4578% 93.75% 4.14% 88.5% 87.7% 87.5% 0.6%45 0.54 0.64 0.75 1.25 ∗ 106 0.4654% 95.31% 4.14% 88.5% 87.7% 87.5% 0.6%46 0.54 0.64 0.75 1.27 ∗ 106 0.4730% 96.88% 4.14% 88.5% 87.7% 87.5% 0.6%47 0.54 0.64 0.75 1.29 ∗ 106 0.4807% 98.44% 4.14% 88.5% 87.7% 87.5% 0.6%48 0.54 0.64 0.75 1.31 ∗ 106 0.4883% 100.00% 4.06% 88.5% 87.7% 87.5% 0.6%* 𝟋 compared to the SML ω256.† 𝟋 compared to the naïve binary SML-CPCA ω256.

Table A.30: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using SMR-CPCA-M templates at T = 64 with session 2 of the PolyUdataset.

163

SMR Workload Recognition Performance

t λmax λCP CAmax MT ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.54 0.63 0.75 6.76 ∗ 105 0.2518% 51.56% 27.50% 71.4% 70.8% 70.5% 17.7%2 0.54 0.64 0.75 6.96 ∗ 105 0.2594% 53.12% 17.66% 80.4% 80.0% 79.5% 8.6%3 0.54 0.64 0.75 7.17 ∗ 105 0.2670% 54.69% 13.28% 84.4% 83.2% 80.6% 7.5%4 0.43 0.61 0.75 7.37 ∗ 105 0.2747% 56.25% 12.50% 84.8% 83.5% 82.6% 5.5%5 0.45 0.63 0.75 7.58 ∗ 105 0.2823% 57.81% 10.08% 86.7% 85.7% 84.2% 3.9%6 0.43 0.61 0.75 7.78 ∗ 105 0.2899% 59.38% 9.38% 87.2% 86.0% 85.1% 3.0%7 0.43 0.60 0.80 7.99 ∗ 105 0.2975% 60.94% 8.44% 87.7% 86.6% 85.7% 2.4%8 0.43 0.60 0.80 8.19 ∗ 105 0.3052% 62.50% 7.58% 88.4% 87.2% 86.2% 1.9%9 0.43 0.61 0.75 8.40 ∗ 105 0.3128% 64.06% 6.41% 89.0% 87.8% 86.8% 1.3%10 0.43 0.61 0.75 8.60 ∗ 105 0.3204% 65.62% 5.78% 89.4% 88.1% 87.1% 1.0%11 0.43 0.61 0.75 8.81 ∗ 105 0.3281% 67.19% 5.47% 89.6% 88.3% 87.3% 0.9%12 0.43 0.61 0.75 9.01 ∗ 105 0.3357% 68.75% 5.16% 89.8% 88.5% 87.5% 0.6%13 0.43 0.61 0.75 9.22 ∗ 105 0.3433% 70.31% 5.08% 89.9% 88.5% 87.5% 0.6%14 0.42 0.61 0.75 9.42 ∗ 105 0.3510% 71.88% 5.08% 89.7% 88.9% 87.7% 0.5%15 0.43 0.62 0.75 9.63 ∗ 105 0.3586% 73.44% 4.69% 90.2% 88.9% 87.7% 0.5%16 0.42 0.60 0.75 9.83 ∗ 105 0.3662% 75.00% 4.84% 90.0% 89.0% 87.7% 0.4%17 0.43 0.61 0.75 1.00 ∗ 106 0.3738% 76.56% 4.14% 90.4% 88.9% 87.9% 0.2%18 0.43 0.61 0.75 1.02 ∗ 106 0.3815% 78.12% 3.91% 90.5% 89.0% 88.0% 0.2%19 0.43 0.61 0.75 1.04 ∗ 106 0.3891% 79.69% 3.83% 90.5% 89.1% 88.0% 0.1%20 0.43 0.61 0.75 1.06 ∗ 106 0.3967% 81.25% 3.75% 90.6% 89.1% 88.1% 0.0%21 0.43 0.61 0.75 1.09 ∗ 106 0.4044% 82.81% 3.75% 90.6% 89.1% 88.1% 0.0%22 0.43 0.61 0.80 1.11 ∗ 106 0.4120% 84.38% 3.83% 90.7% 89.2% 88.2% +0.1%23 0.43 0.61 0.80 1.13 ∗ 106 0.4196% 85.94% 3.91% 90.7% 89.2% 88.2% +0.1%24 0.43 0.61 0.85 1.15 ∗ 106 0.4272% 87.50% 4.38% 89.5% 88.4% 88.1% 0.0%25 0.43 0.61 0.85 1.17 ∗ 106 0.4349% 89.06% 4.38% 89.5% 88.4% 88.1% 0.0%26 0.43 0.61 0.85 1.19 ∗ 106 0.4425% 90.62% 4.30% 89.5% 88.4% 88.1% 0.0%27 0.43 0.61 0.85 1.21 ∗ 106 0.4501% 92.19% 4.14% 89.5% 88.4% 88.2% +0.1%28 0.43 0.61 0.85 1.23 ∗ 106 0.4578% 93.75% 4.14% 89.5% 88.4% 88.2% +0.1%29 0.43 0.61 0.85 1.25 ∗ 106 0.4654% 95.31% 3.91% 89.5% 88.4% 88.2% +0.1%30 0.43 0.61 0.85 1.27 ∗ 106 0.4730% 96.88% 3.83% 89.5% 88.4% 88.2% +0.1%31 0.43 0.61 0.85 1.29 ∗ 106 0.4807% 98.44% 3.98% 89.5% 88.4% 88.2% +0.1%32 0.43 0.61 0.85 1.31 ∗ 106 0.4883% 100.00% 3.83% 89.6% 88.5% 88.3% +0.2%33 0.43 0.61 0.85 1.33 ∗ 106 0.4959% 101.56% 3.83% 89.6% 88.5% 88.3% +0.2%34 0.43 0.61 0.85 1.35 ∗ 106 0.5035% 103.12% 3.98% 89.6% 88.5% 88.3% +0.2%35 0.43 0.61 0.85 1.37 ∗ 106 0.5112% 104.69% 3.83% 89.6% 88.5% 88.3% +0.2%36 0.43 0.61 0.90 1.39 ∗ 106 0.5188% 106.25% 3.75% 89.6% 88.5% 88.3% +0.2%37 0.43 0.61 0.90 1.41 ∗ 106 0.5264% 107.81% 3.75% 89.6% 88.5% 88.3% +0.2%38 0.43 0.61 0.90 1.43 ∗ 106 0.5341% 109.38% 3.75% 89.6% 88.5% 88.3% +0.2%39 0.43 0.61 0.85 1.45 ∗ 106 0.5417% 110.94% 3.91% 89.6% 88.5% 88.3% +0.2%40 0.43 0.61 0.85 1.47 ∗ 106 0.5493% 112.50% 3.91% 89.6% 88.5% 88.3% +0.2%41 0.43 0.61 0.90 1.50 ∗ 106 0.5569% 114.06% 3.75% 89.6% 88.5% 88.3% +0.2%42 0.43 0.61 0.90 1.52 ∗ 106 0.5646% 115.62% 3.75% 89.6% 88.5% 88.3% +0.2%43 0.43 0.61 0.85 1.54 ∗ 106 0.5722% 117.19% 3.91% 89.6% 88.5% 88.3% +0.2%44 0.43 0.61 0.85 1.56 ∗ 106 0.5798% 118.75% 3.91% 89.6% 88.5% 88.3% +0.2%45 0.43 0.61 0.85 1.58 ∗ 106 0.5875% 120.31% 3.91% 89.6% 88.5% 88.3% +0.2%46 0.43 0.61 0.85 1.60 ∗ 106 0.5951% 121.88% 3.91% 89.6% 88.5% 88.3% +0.2%47 0.43 0.61 0.85 1.62 ∗ 106 0.6027% 123.44% 3.91% 89.6% 88.5% 88.3% +0.2%48 0.43 0.61 0.85 1.64 ∗ 106 0.6104% 125.00% 3.91% 89.6% 88.5% 88.3% +0.2%* 𝟋 compared to the SML ω256.† 𝟋 compared to the naïve binary SML-CPCA ω256.

Table A.31: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using SMR-CPCA-M templates at T = 64 with an additional realPSML-CPCA comparison for session 2 of the PolyU dataset.

164

SMR Workload Recognition Performance

t λmax λCP CAmax MT Bit Threshold ω256 𝟋B

* 𝟋R† PER TP0.5 TP0.1 TP0

TP0 loss toPSML-Baseline

1 0.53 0.61 0.75 0.7 3.48 ∗ 105 0.1297% 26.56% 24.06% 74.1% 73.9% 73.6% 16.8%2 0.53 0.60 0.75 0.7 3.69 ∗ 105 0.1373% 28.12% 15.00% 82.7% 81.7% 81.6% 8.8%3 0.48 0.60 0.75 0.5 3.89 ∗ 105 0.1450% 29.69% 12.27% 84.9% 84.5% 84.0% 6.4%4 0.48 0.60 0.75 0.5 4.10 ∗ 105 0.1526% 31.25% 10.23% 86.1% 85.5% 85.1% 5.3%5 0.43 0.60 0.75 0.5 4.30 ∗ 105 0.1602% 32.81% 9.30% 87.4% 86.3% 86.2% 4.2%6 0.43 0.60 0.80 0.6 4.51 ∗ 105 0.1678% 34.38% 7.58% 88.0% 87.3% 87.0% 3.4%7 0.43 0.60 0.75 0.7 4.71 ∗ 105 0.1755% 35.94% 6.72% 89.2% 87.9% 87.8% 2.6%8 0.43 0.60 0.75 0.6 4.92 ∗ 105 0.1831% 37.50% 6.02% 89.5% 88.2% 88.0% 2.4%9 0.43 0.60 0.75 0.7 5.12 ∗ 105 0.1907% 39.06% 5.55% 89.9% 88.4% 88.4% 2.0%10 0.43 0.60 0.75 0.7 5.32 ∗ 105 0.1984% 40.62% 5.39% 89.4% 88.5% 88.4% 2.0%11 0.43 0.60 0.75 0.7 5.53 ∗ 105 0.2060% 42.19% 4.92% 89.8% 88.9% 88.8% 1.6%12 0.43 0.60 0.75 0.7 5.73 ∗ 105 0.2136% 43.75% 4.69% 89.8% 89.0% 88.9% 1.5%13 0.43 0.60 0.75 0.6 5.94 ∗ 105 0.2213% 45.31% 4.61% 89.6% 89.0% 89.0% 1.4%14 0.43 0.60 0.75 0.6 6.14 ∗ 105 0.2289% 46.88% 4.53% 89.6% 89.0% 89.0% 1.4%15 0.43 0.60 0.75 0.6 6.35 ∗ 105 0.2365% 48.44% 4.22% 89.8% 89.2% 89.2% 1.2%16 0.43 0.60 0.75 0.6 6.55 ∗ 105 0.2441% 50.00% 4.30% 89.8% 89.2% 89.2% 1.2%17 0.43 0.60 0.75 0.6 6.76 ∗ 105 0.2518% 51.56% 4.14% 89.8% 89.2% 89.2% 1.2%18 0.43 0.60 0.75 0.6 6.96 ∗ 105 0.2594% 53.12% 4.06% 89.8% 89.2% 89.2% 1.2%19 0.43 0.60 0.75 0.6 7.17 ∗ 105 0.2670% 54.69% 3.83% 89.9% 89.3% 89.3% 1.1%20 0.43 0.60 0.75 0.6 7.37 ∗ 105 0.2747% 56.25% 3.59% 90.1% 89.4% 89.4% 1.0%21 0.43 0.60 0.75 0.6 7.58 ∗ 105 0.2823% 57.81% 3.36% 90.3% 89.5% 89.5% 0.9%22 0.43 0.60 0.75 0.6 7.78 ∗ 105 0.2899% 59.38% 3.36% 90.3% 89.5% 89.5% 0.9%23 0.43 0.60 0.75 0.6 7.99 ∗ 105 0.2975% 60.94% 3.36% 90.3% 89.5% 89.5% 0.9%24 0.43 0.60 0.75 0.6 8.19 ∗ 105 0.3052% 62.50% 3.36% 90.3% 89.5% 89.5% 0.9%25 0.43 0.60 0.75 0.6 8.40 ∗ 105 0.3128% 64.06% 3.36% 90.3% 89.5% 89.5% 0.9%26 0.43 0.60 0.75 0.6 8.60 ∗ 105 0.3204% 65.62% 3.28% 90.4% 89.6% 89.6% 0.8%27 0.43 0.60 0.75 0.6 8.81 ∗ 105 0.3281% 67.19% 3.36% 90.4% 89.6% 89.6% 0.8%28 0.43 0.60 0.75 0.6 9.01 ∗ 105 0.3357% 68.75% 3.36% 90.4% 89.6% 89.6% 0.8%29 0.43 0.60 0.75 0.6 9.22 ∗ 105 0.3433% 70.31% 3.36% 90.4% 89.6% 89.6% 0.8%30 0.43 0.60 0.75 0.6 9.42 ∗ 105 0.3510% 71.88% 3.36% 90.4% 89.6% 89.6% 0.8%31 0.43 0.60 0.75 0.6 9.63 ∗ 105 0.3586% 73.44% 3.28% 90.5% 89.6% 89.6% 0.8%32 0.43 0.60 0.75 0.6 9.83 ∗ 105 0.3662% 75.00% 3.28% 90.5% 89.6% 89.6% 0.8%33 0.43 0.60 0.75 0.6 1.00 ∗ 106 0.3738% 76.56% 3.28% 90.5% 89.6% 89.6% 0.8%34 0.43 0.60 0.75 0.6 1.02 ∗ 106 0.3815% 78.12% 3.28% 90.5% 89.6% 89.6% 0.8%35 0.43 0.60 0.75 0.6 1.04 ∗ 106 0.3891% 79.69% 3.28% 90.5% 89.6% 89.6% 0.8%36 0.43 0.60 0.75 0.6 1.06 ∗ 106 0.3967% 81.25% 3.28% 90.2% 89.6% 89.6% 0.8%37 0.43 0.60 0.75 0.6 1.09 ∗ 106 0.4044% 82.81% 3.28% 90.2% 89.6% 89.6% 0.8%38 0.43 0.60 0.75 0.6 1.11 ∗ 106 0.4120% 84.38% 3.28% 90.2% 89.6% 89.6% 0.8%39 0.43 0.60 0.75 0.6 1.13 ∗ 106 0.4196% 85.94% 3.28% 90.2% 89.6% 89.6% 0.8%40 0.43 0.60 0.75 0.6 1.15 ∗ 106 0.4272% 87.50% 3.28% 90.2% 89.6% 89.6% 0.8%41 0.43 0.60 0.75 0.6 1.17 ∗ 106 0.4349% 89.06% 3.20% 90.2% 89.6% 89.6% 0.8%42 0.43 0.60 0.75 0.6 1.19 ∗ 106 0.4425% 90.62% 3.20% 90.2% 89.6% 89.6% 0.8%43 0.43 0.60 0.75 0.6 1.21 ∗ 106 0.4501% 92.19% 3.20% 90.2% 89.6% 89.6% 0.8%44 0.43 0.60 0.75 0.6 1.23 ∗ 106 0.4578% 93.75% 3.20% 90.2% 89.6% 89.6% 0.8%45 0.43 0.60 0.75 0.6 1.25 ∗ 106 0.4654% 95.31% 3.20% 90.2% 89.6% 89.6% 0.8%46 0.43 0.60 0.75 0.6 1.27 ∗ 106 0.4730% 96.88% 3.20% 90.2% 89.6% 89.6% 0.8%47 0.43 0.60 0.75 0.6 1.29 ∗ 106 0.4807% 98.44% 3.12% 90.2% 89.6% 89.6% 0.8%48 0.43 0.60 0.75 0.6 1.31 ∗ 106 0.4883% 100.00% 3.12% 90.2% 89.6% 89.6% 0.8%* 𝟋 compared to the SML ω256.† 𝟋 compared to the naïve binary SML-CPCA ω256.

Table A.32: Best achieved CPCA-Tree TP0 with corresponding TP0.1 and TP0.5 for every tree pre-selection of t = 1, . . . , 48 trees using SMR-CPCA-M templates at T = 64 with an additional binarySML-CPCA comparison and masking out most common set bits for session 2 of the PolyU dataset.

165

Name δ(TP0,𝟋B) 𝟋B* TP0

Bf CPCA ARC 0.10 0.1952 89.0%C-T SCM Masking 0.14 0.2384 89.0%C-T SCM 0.14 0.2441 89.0%C-T SCM ARC 0.29 0.3128 88.8%Binary PSML-CPCA† 0.39 0.4880 89.0Bf CPCA ABC 1.03 0.3437 88.0%

Name δ(TP0.1,𝟋B) 𝟋B* TP0.1

Bf CPCA ARC 0.10 0.1952 89.1%C-T SCM Masking 0.12 0.2155 89.1%C-T SCM 0.14 0.2365 89.1%C-T SCM ARC 0.29 0.3052 88.9%Binary PSML-CPCA† 0.39 0.4880 89.1%Bf CPCA ABC 0.32 0.3437 88.9%* 𝟋 compared to the PSML ω256.† Naïve approach.

Table A.33: Ranking for each relevant experiment, respectively workload reduction method, orderedby the euclidean distance rating method with optimal operation point at naïve binary PSML-CPCAbiometric performance (TP0 = 89.0) and 𝟋 = 0.1%.

166

List of Symbols

BC - Bloom filter set size Amount of Bloom filter Bin a set B. 46, 48

MCP CA - SMR height Height of an SMR-CPCA template. 58–60, 65, 90, 91, 94

M - SMR height Height of an SMR template. 46, 60, 65

N - SMR width Width of an SMR template. 46, 65, 94

PER - Pre-selection Error Rate Pre-selection error: error that occurs when the corres-ponding enrolment template is not in the pre-selected subset of candidates when a samplefrom the same biometric characteristic on the same user is given; PER: rate of falsely se-lected rank N (here 1) candidates. 94, 105–111, 127, 147–155, 158–160, 163–165

𝟋 - Workload Fraction Workload reduction expressed as fraction of the baseline workload(see section 8.5). 75, 76, 90–93, 95–105, 107, 108, 110, 111, 113–118, 136–166

B - Bloom filter A single Bloom filter. 45, 46, 48, 53, 54, 167, 168

C - Bloom filter-Trees Amount of template comparisons of a single lookup in an identifica-tion scenario. 51–53

H - Bloom filter height Block-size height used for the Bloom filter template transformation.46, 49, 50, 53, 54, 93–105, 136–146, 157, 162

S - Subjects Enrolled subjects in database. Note: one subject is able to present two instances.Since there is no possibility to link two instances to one subject, one instance is treatedas one subject in this thesis. 10, 47, 48, 50–53, 62, 75, 76, 88, 119, 126, 128

T - Bloom filter-Trees Amount of constructed trees in a Bloom filter identification scenario.50, 51, 60, 61, 93–102, 105–111, 124, 136–141, 143–155, 157–160, 162–165

W - Bloom filter width Block-size width used for the Bloom filter template transformation.46, 49, 53, 54, 93–105, 136–146, 157, 162

X - SMR Spectrum Spectrum symbol for the SMR. 32–35, 65

L - CPCA training set size Amount of used SMR templates for the training of the CPCA(see section 5.4.1). 89, 134

167

MT - Mask Threshold Threshold value used during binarisation of the SMR spectrum. 34,35, 90–93, 95–105, 107, 108, 110, 111, 114, 136–165

t - Selected Bloom filter-Trees Amount of selected trees during a tree pre-selection step.51, 52, 96–98, 100–102, 106–108, 110, 111, 114, 116, 117, 119, 138–140, 144–160, 162–165

B - Bloom filter set A set of Bloom filter B. 47, 48, 167

168

List of Abbreviations

CPCA - Column Principal Component Analsysis See section 5.4.1. 31–34, 56–67, 72,74, 75, 89–91, 134

CPCA-Tree - SMR-CPCA binary search tree Binary search tree build with SMR-CPCAtemplates. See chapter 7. 59–61, 67, 74, 105–111, 113–116, 119, 126–128, 147–155, 158–160, 163–165

FHD - fractional Hamming Distance Algorithm used to compare two masked binary feature-vectors (see 5.6.2). 60

LDFT - Line Discrete Fourier Transformation See section 5.4.2. 33

MHD - modified Hausdorf Distance Algorithm used to compare feature-vectors (set ofpoints, see 9.1). 77–84, 86, 87, 120

NIR - Near-Infrared Radiation with wavelength from about 700 nm to 2500 nm, thus invis-ible to the human eye. 4–6, 8, 12, 19, 20, 23, 68, 70

NL-diffusion - Non-Linear Diffusion Image enhancement algorithm from [Wei01]. Used toreduce high-frequency noise and enhancing contrast. 23, 24, 77

NL-means - Non-Local Means Image enhancement algorithm from [ŠP09]. Used to createan illumination invariant texture by non-local smoothing. 22–24, 77

PSMC - Spectral Minutia Complex Representation with minutiae pre-selection Seesection 5.8. 85

PSML - Spectral Minutia Location Representation with minutiae pre-selection Seesection 5.8. 75, 86, 88–93, 95–105, 107, 108, 110–115, 119, 136–155, 166

PSML-CPCA - PSML reduced with the CPCA feature-reduction See section 5.4.1.88–111, 113–116, 136–142, 147–155, 164, 166

PSML-CPCA-A - PSML-CPCA-Applied PSML-CPCA representation with applied mask-bit on the sign-bit. See section 7.1.4. 106, 108, 109

169

PSML-CPCA-C - PSML-CPCA-Components The sign-bit and mask-bit of the PSML-CPCA. see section 7.1.4. 106–108

PSML-CPCA-M - PSML-CPCA-Mixed Same as PSML-CPCA-Applied but keeping themask-bit. See section 7.1.4. 106–111, 114, 117

PSMR - Spectral Minutia Representation with minutiae pre-selection See section 5.8.42, 82, 85, 87

QSMC - quality data enhanced Spectral Minutia Complex Representation See sec-tion 5.7. 42, 82, 83, 86, 122

QSML - quality data enhanced Spectral Minutia Location Representation See sec-tion 5.7. 38, 42, 81–83, 86, 88, 121, 134

ROC - Receiver Operating Characteristic Curve Plot of the rate of false positives (i.e.impostor attempts accepted) against the corresponding rate of true positives as a functionof the decision threshold. 80–91, 112, 116, 118, 121, 122, 133–135

ROI - Region of Interest Part of an image which contains the desired informations. 8, 18–24, 68–72, 76, 77, 83, 85, 121

SMC - Spectral Minutia Complex Representation See section 5.1.3. 29, 30, 32, 42, 65,81–86, 122

SML - Spectral Minutia Location Representation See section 5.1.1. 28–30, 32, 34, 36,37, 42, 54, 57, 65, 66, 81–83, 86, 89, 112, 113, 121, 122, 134, 156–165

SML-CPCA - SML reduced with the CPCA feature-reduction See section 5.4.1. 57,58, 155–165

SMO - Spectral Minutia Orientation Representation See section 5.1.2. 29, 32, 81

SMR - Spectral Minutia Representation See section 5.1. 27–38, 40, 42, 45, 46, 49, 53–57, 59, 61, 63–67, 73, 74, 76, 80–93, 95–105, 107, 108, 110–113, 116, 119, 121, 123–129,134–165

SMR-CPCA - SMR reduced with the CPCA feature-reduction See section 5.4.1. 36,56–61, 64, 65, 74, 90, 106, 116, 119, 125–128

SMR-CPCA-A - SMR-CPCA-Applied SMR-CPCA representation with applied mask-bit on the sign-bit. See section 7.1.4. 59, 60, 106

SMR-CPCA-C - SMR-CPCA-Components The sign-bit and mask-bit of the SMR-CPCA.see section 7.1.4. 59, 60, 106, 147

170

SMR-CPCA-M - SMR-CPCA-Mixed Same as SMR-CPCA-Applied but keeping the mask-bit. See section 7.1.4. 60, 106, 148–156, 158–160, 163–165

171

Bibliography

[AG14] AM Abbas and Loay E George, Palm vein recognition and verification systemusing local average of vein direction, Int. J. Sci. Eng. Res 5 (2014), no. 4, 1026–1033.

[Bad06] Ahmed M Badawi, Hand vein biometric verification prototype: A testing perform-ance and patterns similarity., IPCV 14 (2006), 3–9.

[BLM14] Richard Barnes, Clarence Lehman, and David Mulla, Priority-flood: An optimaldepression-filling and watershed-labeling algorithm for digital elevation models,Computers & Geosciences 62 (2014), 117–127.

[Blo70] Burton H Bloom, Space/time trade-offs in hash coding with allowable errors,Communications of the ACM 13 (1970), no. 7, 422–426.

[BP98] Ruud Bolle and Sharath Pankanti, Biometrics, personal identification in net-worked society: Personal identification in networked society, Kluwer AcademicPublishers, Norwell, MA, USA, 1998.

[Bus10] Christoph Busch, Iso 24745-biometric template protection, The first InternationalBiometric Performance Testing Conference, Match, 2010.

[cas] Casia-ms-palmprintv1, http://biometrics.idealtest.org/.

[CDG+08] Fay Chang, Jeffrey Dean, Sanjay Ghemawat, Wilson C Hsieh, Deborah A Wal-lach, Mike Burrows, Tushar Chandra, Andrew Fikes, and Robert E Gruber, Bigt-able: A distributed storage system for structured data, ACM Transactions onComputer Systems (TOCS) 26 (2008), no. 2, 4.

[CFM10] Raffaele Cappelli, Matteo Ferrara, and Davide Maltoni, Minutia cylinder-code:A new representation and matching technique for fingerprint recognition, IEEETransactions on Pattern Analysis and Machine Intelligence 32 (2010), no. 12,2128–2141.

[CFM11] , Fingerprint indexing based on minutia cylinder-code, IEEE Transactionson Pattern Analysis and Machine Intelligence 33 (2011), no. 5, 1051–1057.

172

[CP76] David Casasent and Demetri Psaltis, Position, rotation, and scale invariant op-tical correlation, Applied optics 15 (1976), no. 7, 1795–1799.

[Cre16] Credence Research, Biometrics technology market by technology (face recognition,hand geometry, voice recognition, signature recognition, iris recognition, afis, non-afis & others), end use vertical (government, military and defense, healthcare,banking and finance, consumer electronics, residential and commercial security,transportation and immigration & others) - growth, share, opportunities & com-petitive analysis, 2015 - 2022, 4 2016.

[csp13a] Case study: Palmsecure - maximum access security for sensit-ive areas, https://sp.ts.fujitsu.com/dmsp/Publications/public/cs-PalmSecure-PCS-TDS-en.pdf, 2013.

[csp13b] Case study: Palmsecure - over 700 million transactions without any fraudincident, http://www.fujitsu.com/global/Images/cs-banco-bradesco-en.pdf, 2013.

[csp14] Case study: Palmsecure - ferencváros soccer club scores with biosec biometricpersonal identification system for its new stadium, https://sp.ts.fujitsu.com/dmsp/Publications/public/cs-PalmSecure-BioSec.pdf, 2014.

[Dau00] John Daugman, Biometric decision landscapes, Tech. report, University of Cam-bridge, Computer Laboratory, 2000.

[Dau03] , The importance of being random: statistical principles of iris recognition,Pattern recognition 36 (2003), no. 2, 279–291.

[Dau04] , How iris recognition works, IEEE Transactions on circuits and systemsfor video technology 14 (2004), no. 1, 21–30.

[DJ94] M.-P. Dubuisson and A.K. Jain, A modified hausdorff distance for object match-ing, Proceedings of 12th International Conference on Pattern Recognition, IEEEComput. Soc. Press, 1994.

[DRB17] Pawel Drozdowski, Christian Rathgeb, and Christoph Busch, Bloom filter-basedsearch structures for indexing and retrieving iris-codes, IET Biometrics (2017).

[FZ05] Marcos Faundez-Zanuy, Privacy issues on biomet-ric systems, IEEE A&E SYSTEMS MAGAZINE (2005),http://ai.pku.edu.cn/aiwebsite/research.files/collected%20papers%20-%20introduction/Privacy%20issues%20on%20biometric%20systems.pdf.

173

[GBRG+16] Marta Gomez-Barrero, Christian Rathgeb, Javier Galbally, Christoph Busch, andJulian Fierrez, Unlinkable and irreversible biometric template protection based onbloom filters, Information Sciences 370 (2016), 18–32.

[GH89] Zicheng Guo and Richard W Hall, Parallel thinning with two-subiteration al-gorithms, Communications of the ACM 32 (1989), no. 3, 359–373.

[GL18] H. Gray and W.H. Lewis, Anatomy of the human body, Lea & Febiger, 1918.

[GY83] Ronald L Graham and F Frances Yao, Finding the convex hull of a simple polygon,Journal of Algorithms 4 (1983), no. 4, 324–331.

[GZZ+11] Zhenhua Guo, David Zhang, Lei Zhang, Wangmeng Zuo, and Guangming Lu,Empirical study of light source selection for palmprint recognition, Pattern Re-cognition Letters 32 (2011), no. 2, 120–126.

[Har12] Daniel Hartung, Vascular pattern recognition and its application in privacy-preserving biometric online-banking systems, Ph.D. thesis, Gjovik University Col-lege, Faculty of Computer Science and Media Technology, 2 2012.

[HDZ08] Feng Hao, John Daugman, and Piotr Zielinski, A fast search algorithm for alarge fuzzy database, IEEE Transactions on Information Forensics and Security 3(2008), no. 2, 203–212.

[Hes12] Scott Hess, Issue 10896048: Transition safe browsing from bloom filter to prefixset., 08 2012.

[HGZ08] Dong Han, Zhenhua Guo, and David Zhang, Multispectral palmprint recognitionusing wavelet-based image fusion, Signal Processing, 2008. ICSP 2008. 9th Inter-national Conference on, IEEE, 2008, pp. 2074–2077.

[HOX+12] Daniel Hartung, Martin Aastrup Olsen, Haiyun Xu, Hai Thanh Nguyen, andChristoph Busch, Comprehensive analysis of spectral minutiae for vein patternrecognition., IET Biometrics 1 (2012), no. 1, 25–36.

[HOXB11] Daniel Hartung, Martin Aastrup Olsen, Haiyun Xu, and Christoph Busch, Spec-tral minutiae for vein pattern recognition, Biometrics (IJCB), 2011 InternationalJoint Conference on, IEEE, 2011, pp. 1–7.

[Ind] India, Aadhaar issued summary.

[iso06] ISO/IEC 19795-1:2006 - Information technology – Biometric performance testingand reporting – Part 1: Principles and framework, International Standard, 2006.

174

[JHF14] Y. Jo, S. Hama, and M. Fukuda, Vein authentication method, image pro-cessing method, and vein authentication device, October 2 2014, US Patent App.14/299,435.

[JHZ08] Wei Jia, De-Shuang Huang, and David Zhang, Palmprint verification based onrobust line orientation code, Pattern Recognition 41 (2008), no. 5, 1504–1513.

[KK10] Rafał Kabaciński and Mateusz Kowalski, Human vein pattern segmentation fromlow quality images–a comparison of methods, Image Processing and Communic-ations Challenges 2 (2010), no. 84, 105–112.

[KK11a] R Kabacinski and Mateusz Kowalski, Vein pattern database and benchmark res-ults, Electronics Letters 47 (2011), no. 20, 1127–1128.

[KK11b] Rafał Kabacinski and Mateusz Kowalski, Human vein pattern correlation-a com-parison of segmentation methods, Computer Recognition Systems 4, 95 of Ad-vances in Intelligent and Soft Computing (2011), 51–59.

[Koc13] Davit Kocharyan, A modified fingerprint image thinning algorithm, AmericanJournal of Software Engineering and Applications 2 (2013), no. 1, 1–6.

[Llo82] Stuart Lloyd, Least squares quantization in pcm, IEEE transactions on informa-tion theory 28 (1982), no. 2, 129–137.

[LT15] Sen Lin and Yu Tai, A combination recognition method of palmprint and palmvein based on gray surface matching, Image and Signal Processing (CISP), 20158th International Congress on, IEEE, 2015, pp. 567–571.

[Lun61] RJ Lunnon, Some observations on the photography of the diseased skin., Medical& biological illustration 11 (1961), 98.

[MNM04] Naoto Miura, Akio Nagasaka, and Takafumi Miyatake, Feature extraction offinger-vein patterns based on repeated line tracking and its application to personalidentification, Machine Vision and Applications 15 (2004), no. 4, 194–203.

[MNM07] , Extraction of finger-vein patterns using maximum curvature points inimage profiles., IEICE Transactions 90-D (2007), no. 8, 1185–1194.

[MS15] Bruce M Maggs and Ramesh K Sitaraman, Algorithmic nuggets in content de-livery, ACM SIGCOMM Computer Communication Review 45 (2015), no. 3,52–66.

[Nad07] Annemarie Nadort, The hand vein pattern used as a biometric feature, Master’sthesis, Free University, Amsterdam, 3 2007.

175

[OHBL11] Martin Aastrup Olsen, Daniel Hartung, Christoph Busch, and Rasmus Larsen,Convolution approach for feature detection in topological skeletons obtained fromvascular patterns, 2011 IEEE Workshop on Computational Intelligence in Bio-metrics and Identity Management (CIBIM), IEEE, apr 2011.

[OWN96] Alan V Oppenheim, Alan S Willsky, and S Hamid Nawab, Signals & Systems,Prentice-Hall, Inc., Upper Saddle River, NJ, USA 18 (1996), 19–21.

[Pea01] Karl Pearson, On lines and planes of closest fit to systems of points in space, TheLondon, Edinburgh, and Dublin Philosophical Magazine and Journal of Science2 (1901), no. 11, 559–572.

[PGS+09] A. Jayaprakash Patil, Ana L. Gramajo, Ashish Sharma, Gail M. Seigel, Baruch D.Kuppermann, and M. Cristina Kenney, Differential effects of nicotine on retinaland vascular cells in vitro, Toxicology 259 (2009), no. 1-2, 69–76.

[Pit00] Michael Pittilo, Cigarette smoking, endothelial injury and cardiovascular disease,International Journal of Experimental Pathology 81 (2000), no. 4, 219–230.

[PK11] Mi Pan and Wenxiong Kang, Palm vein recognition based on three local invari-ant feature extraction algorithms, Chinese Conference on Biometric Recognition,Springer, 2011, pp. 116–124.

[pol] Polyu multispectral palmprint database, http://www.comp.polyu.edu.hk/~biometrics/MultispectralPalmprint/MSP.htm.

[Pro13] Hugo Proenca, Iris biometrics: Indexing and retrieving heavily degraded data,IEEE Transactions on Information Forensics and Security 8 (2013), no. 12, 1975–1985.

[PUASRL10] J Enrique Suarez Pascual, Jaime Uriarte-Antonio, Raul Sanchez-Reillo, and Mi-chael G Lorenz, Capturing hand or wrist vein images for biometric authentica-tion using low-cost devices, Intelligent Information Hiding and Multimedia Sig-nal Processing (IIH-MSP), 2010 Sixth International Conference on, IEEE, 2010,pp. 318–322.

[RB13] Christian Rathgeb and Christoph Busch, Comparing binary iris biometric tem-plates based on counting bloom filters., CIARP (2) (José Ruiz-Shulcloper andGabriella Sanniti di Baja, eds.), Lecture Notes in Computer Science, vol. 8259,Springer, 2013, pp. 262–269.

[RBB13] Christian Rathgeb, Frank Breitinger, and Christoph Busch, Alignment-free can-celable iris biometric templates based on adaptive bloom filters., ICB (Julian Fiér-rez, Ajay Kumar, Mayank Vatsa, Raymond N. J. Veldhuis, and Javier Ortega-Garcia, eds.), IEEE, 2013, pp. 1–8.

176

[RBBB15] Christian Rathgeb, Harald Baier, Christoph Busch, and Frank Breitinger, To-wards bloom filter-based indexing of iris biometric data., ICB, IEEE, 2015,pp. 422–429.

[Ric85] J. Rice, Method and apparatus for the identification of individuals, September 261985, WO Patent App. PCT/GB1985/000,127.

[RJ+91] Sarunas J Raudys, Anil K Jain, et al., Small sample size effects in statisticalpattern recognition: Recommendations for practitioners, IEEE Transactions onpattern analysis and machine intelligence 13 (1991), no. 3, 252–264.

[RU11] Christian Rathgeb and Andreas Uhl, A survey on biometric cryptosystems andcancelable biometrics, EURASIP Journal on Information Security 2011 (2011),no. 1, 3.

[S+85] Satoshi Suzuki et al., Topological structural analysis of digitized binary imagesby border following, Computer vision, graphics, and image processing 30 (1985),no. 1, 32–46.

[Skl82] Jack Sklansky, Finding the convex hull of a simple polygon, Pattern RecognitionLetters 1 (1982), no. 2, 79–83.

[SKS+14] T. Shindo, T. Kamimura, N. Suzuki, K. Ogawa, S. Eguchi, and M. MANABE,Authentication apparatus, image capture apparatus, authentication method, andauthentication program, February 11 2014, US Patent 8,649,569.

[ŠP09] Vitomir Štruc and Nikola Pavešić, Illumination invariant face recognition by non-local smoothing, European Workshop on Biometrics and Identity Management,Springer, 2009, pp. 1–8.

[SRB15] Jayachander Surbiryala, Ramachandra Raghavendra, and Christoph Busch, Fin-ger vein indexing based on binary features, Colour and Visual Computing Sym-posium (CVCS), 2015, IEEE, 2015, pp. 1–6.

[The06] M Theme, Comparative biometric testing round 6 public report, The InternationalBiometric Group (2006).

[Tra14] Transparency Market Research, Biometrics technology (face, hand geometry,voice, signature, iris, afis, non-afis and others) market - global industry analysissize share growth trends and forecast 2013 - 2019, 4 2014.

[Tra17] Tractica LLC, Biometrics market forecasts: Global unit shipments and revenueby biometric modality, technology, use case, industry segment, and world region:2016-2025, Q1 2017.

177

[Wat14] M. Watanabe, Biometric authentication apparatus and biometric authenticationmethod, August 19 2014, US Patent 8,811,681.

[Wei01] Joachim Weickert, Applications of nonlinear diffusion in image processing andcomputer vision, Acta Math. Univ. Comenianae 70 (2001), no. 1, 33–50.

[WESS05] Masaki Watanabe, Toshio Endoh, Morito Shiohara, and Shigeru Sasaki, Palmvein authentication technology and its applications, Proceedings of the biometricconsortium conference, 2005, pp. 19–21.

[WL06] Lingyu Wang and Graham Leedham, Near-and far-infrared imaging for veinpattern biometrics, Video and Signal Based Surveillance, 2006. AVSS’06. IEEEInternational Conference on, IEEE, 2006, pp. 52–52.

[XV09] Haiyun Xu and Raymond NJ Veldhuis, Spectral minutiae representations of fin-gerprints enhanced by quality data, Biometrics: Theory, Applications, and Sys-tems, 2009. BTAS’09. IEEE 3rd International Conference on, IEEE, 2009, pp. 1–5.

[XV10a] , Binary representations of fingerprint spectral minutiae features, Pat-tern Recognition (ICPR), 2010 20th International Conference on, IEEE, 2010,pp. 1212–1216.

[XV10b] , Complex spectral minutiae representation for fingerprint recognition,Computer Vision and Pattern Recognition Workshops (CVPRW), 2010 IEEEComputer Society Conference on, IEEE, 2010, pp. 1–8.

[XV10c] , Spectral minutiae representations for fingerprint recognition, IntelligentInformation Hiding and Multimedia Signal Processing (IIH-MSP), 2010 SixthInternational Conference on, IEEE, 2010, pp. 341–345.

[XVB+09] Haiyun Xu, Raymond NJ Veldhuis, Asker M Bazen, Tom AM Kevenaar,Ton AHM Akkermans, and Berk Gokberk, Fingerprint verification using spec-tral minutiae representations, IEEE Transactions on Information Forensics andSecurity 4 (2009), no. 3, 397–409.

[XVK+08] Haiyun Xu, Raymond NJ Veldhuis, Tom AM Kevenaar, Anton HM Akkermans,and Asker M Bazen, Spectral minutiae: A fixed-length representation of a minutiaeset, Computer Vision and Pattern Recognition Workshops, 2008. CVPRW’08.IEEE Computer Society Conference on, IEEE, 2008, pp. 1–6.

[XVKA09] Haiyun Xu, Raymond NJ Veldhuis, Tom AM Kevenaar, and Ton AHM Akker-mans, A fast minutiae-based fingerprint recognition system, IEEE Systems journal3 (2009), no. 4, 418–427.

178

[Yak10] Alex Yakunin, Nice bloom filter application, 03 2010.

[Zae11] Naser Zaeri, Minutiae-based fingerprint extraction and recognition, INTECHOpen Access Publisher, 2011.

[ZGL+10] David Zhang, Zhenhua Guo, Guangming Lu, Lei Zhang, and Wangmeng Zuo,An online system of multispectral palmprint verification, IEEE transactions oninstrumentation and measurement 59 (2010), no. 2, 480–490.

[ZK11] Yingbo Zhou and Ajay Kumar, Human identification using palm-vein images,IEEE transactions on information forensics and security 6 (2011), no. 4, 1259–1274.

[ZKYW03] David Zhang, Wai-Kin Kong, Jane You, and Michael Wong, Online palmprintidentification, IEEE Transactions on pattern analysis and machine intelligence25 (2003), no. 9, 1041–1050.

[ZLF+14] Y Zhou, Y Liu, Q Feng, F Yang, J Huang, Ian McLoughlin, et al., Palm-veinclassification based on principal orientation features, PLoS ONE 9 (2014), no. 11,e112429.

I